Enhancing user interfaces using generative models

Abstract

A method implemented by one or more computing devices comprises receiving voice input that conveys a user instruction with reference to a displayed user interface. The method includes monitoring a location of a point of attention of the user on the displayed user interface and identifying one or more elements of the displayed user interface based on the location of the point of attention at one or more times. The method further comprises forming a prompt for a generative model based on the user input and the one or more elements so as to generate computer code based on the prompt.

Description

ENHANCING USER INTERFACES USING GENERATIVE MODELSFIELD

[0001] This specification relates to enhancing user interfaces using generative models. In particular, it relates to forming a prompt for a generative model based on user input so as to generate computer code for carrying out a user instruction.BACKGROUND

[0002] Computer user interfaces have evolved significantly since the early days of computing. Starting with basic command-line systems, user interfaces have developed to become more visual and interactive. The introduction of the computer mouse enabled users to point and click, leading to widespread adoption of displayed user interfaces including windows, icons, menus etc. While modern user interfaces may also include touchscreens, voice control, gesture recognition and other features, they still face limitations in bridging the gap between human intent and the predetermined protocols of computing systems.SUMMARY

[0003] Aspects and advantages of embodiments of the present disclosure will be set forth in part in the following description, or can be learned from the description, or can be learned through practice of the embodiments.

[0004] In one aspect, the present disclosure provides a first example method which is implemented by one or more computing devices. The first example method includes receiving voice input that conveys a user instruction with reference to a displayed user interface. The first example method includes monitoring a location of a point of attention of the user on the displayed user interface. The first example method includes identifying one or more elements of the displayed user interface based on the location of the point of attention at one or more times. The first example method includes forming a prompt for a generative model based on the voice input and the one or more elements. The first example method includes obtaining computer code, comprising sending the prompt to the generative model, and executing the computer code.

[0005] In one aspect, the present disclosure provides a second example method which is implemented by one or more computing devices. The second example method includes obtaininga context stream comprising a stream of context data over time. The second example method includes receiving user input that refers to the context stream. The second example method includes evaluating the user input for one or more trigger events. The second example method includes determining one or more prompt items by sampling from the context stream in temporal correspondence with the one or more trigger events, and forming a prompt for a generative model using the one or more prompt items.

[0006] In one aspect, the present disclosure provides a third example method implemented by one or more computing devices. The third example method includes monitoring a location of a point of attention of a user on a displayed user interface. The third example method includes determining one or more prompt items, wherein the one or more prompt items include at least one prompt item which is determined based on the location of the point of attention of the user at a respective time, wherein each of the one or more prompt items includes a corresponding timestamp. The third example method includes forming a prompt for a generative model based on the one or more prompt items.

[0007] Particular embodiments of the subject matter described in this specification can be implemented so as to realize one or more of the following advantages.

[0008] While conventional user interfaces (UIs) allow users to perform a wide variety of tasks, there are in general many complex tasks which can only be performed with such UIs either by breaking the task down into multiple sub-tasks, or by writing code to perform the task. As well as being inefficient for the user, breaking a task into UI sub-tasks is also computationally inefficient (in terms of processor cycles and / or memory usage) if there is a more efficient way to perform the task by executing appropriate code. In contrast, various example implementations described in this specification provide an improved control mechanism which allows complex instructions to be performed efficiently by generating appropriate computer code responsive to a user instruction that is provided with reference to a displayed UI. In this way, various example implementations expand on the capabilities of existing user interfaces whilst also improving their computational efficiency.

[0009] Various example implementations provide an input mechanism for integrating a user instruction into a prompt for a generative model such that the prompt explicitly includes or refers to one or more elements of a displayed user interface that the instruction relates to. By identifying one or more elements of the displayed user interface based on the location of a pointof attention of the user at one or more times, and forming the prompt based on user input and the one or more elements, various example implementations provide an input mechanism which improves compatibility with the predetermined protocols of the underlying computing system.

[0010] In an aspect, the present disclosure provides one or more non-transitory computer readable media that store instructions that are executable by one or more processors to cause a computing system to execute any of the methods described herein.

[0011] In an aspect, the present disclosure provides a computing system having one or more processors and one or more non-transitory computer-readable media that store instructions that are executable by the one or more processors to cause the computing system to execute any of the methods described herein.

[0012] Other example aspects of the present disclosure are directed to other systems, methods, apparatuses, tangible non-transitory computer-readable media, and devices for performing functions described herein. These and other features, aspects, and advantages of various implementations will become better understood with reference to the following description and appended claims. The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate implementations of the present disclosure and, together with the description, help explain the related principles.BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 illustrates components of a system in accordance with an example implementation;

[0014] Figures 2(a) and 2(b) show a first example scenario illustrating use of an enhanced interaction mode of the system of Figure 1;

[0015] Figures 3(a), 3(b), 3(c) and 3(d) show a second example scenario illustrating use of the enhanced interaction mode;

[0016] Figures 4(a), 4(b), 4(c), 4(d), 4(e) and 4(f) show a third example scenario illustrating use of the enhanced interaction mode to fill out a form;

[0017] Figures 5(a) and 5(b) show a further example scenario illustrating use of the enhanced interaction mode to fill out a form;

[0018] Figures 6(a) and 6(b) show an example illustrating use of the enhanced interaction mode to change the attributes of displayed elements.

[0019] Figures 7(a), 7(b) and 7(c) shows an example illustrating use of the enhanced interaction mode for image editing;

[0020] Figures 8(a), 8(b) and 8(c) shows an example illustrating use of the enhanced interaction mode for editing charts;

[0021] Figure 9 shows an example illustrating use of the enhanced interaction mode for visual programming;

[0022] Figures 10(a) and 10(b) shows another example illustrating use of the enhanced interaction mode for visual programming;

[0023] Figures 11(a) and 11(b) show an example illustrating use of the enhanced interaction mode to augment code with comments and / or log functionality;

[0024] Figures 12(a), 12(b) and 12(c) show an example illustrating use of the enhanced interaction mode for document editing;

[0025] Figures 13(a) and 13(b) show an example illustrating use of the enhanced interaction mode for editing a spreadsheet;

[0026] Figure 14 shows an example in which a user conveys an instruction using text rather than as a voice command;

[0027] Figures 15, 16, 17 and 18 illustrate methods in accordance with example implementations;

[0028] Figure 19 is a flow chart diagram illustrating an example method for training a machine-learned model according to example implementations of aspects of the present disclosure;

[0029] Figure 20 is a block diagram of an example processing flow for using machine-learned model(s) to process input(s) to generate output(s) according to example implementations of aspects of the present disclosure;

[0030] Figure 21 is a block diagram of an example sequence processing model according to example implementations of aspects of the present disclosure;

[0031] Figure 22 is a block diagram of an example technique for populating an example input sequence for processing by a sequence processing model according to example implementations of aspects of the present disclosure;

[0032] Figure 23 is a block diagram of an example model development platform according to example implementations of aspects of the present disclosure;

[0033] Figure 24 is a block diagram of an example training workflow for training a machine-learned model according to example implementations of aspects of the present disclosure;

[0034] Figure 25 is a block diagram of an inference system for operating one or more machine-learned model(s) to perform inference according to example implementations of aspects of the present disclosure;

[0035] Figure 26 is a block diagram of an example networked computing system according to example implementations of aspects of the present disclosure;

[0036] Figure 27 is a block diagram of an example computing device according to example implementations of aspects of the present disclosure; and

[0037] Figure 28 is a block diagram of an example computing device according to example implementations of aspects of the present disclosure.DETAILED DESCRIPTION

[0038] Figure 1 illustrates components of a system 100 in accordance with an example implementation. The system 100 may be implemented on a desktop computer. However, systems described in this specification may alternatively be implemented on other suitable hardware, e.g., on a user device such as a tablet or smartphone.

[0039] As shown, the system 100 has an operating system 110 which is configured to communicate with various hardware components including any peripheral devices. The operating system 110 manages a set of device drivers 120 for various hardware components such as keyboard 122, mouse 124, microphone 126, display screen 128 and network interface device 129. In some examples, the set of device drivers 120 may also include a touch input driver 125. The device drivers 120 interface with their respective hardware components and make the resulting data available for processing by the system 100. It will be understood that the set of hardware components illustrated in Figure 1 are merely exemplary and different sets of components may be present in different examples. For instance, systems implemented on certain devices such as smartphones or tablets may include a touchscreen for user input rather than a keyboard and mouse.

[0040] The operating system 110 acts as a platform on which various applications and software that is supported by the system 100 can run. It also provides a displayed user interface to enable users to interact with and control the system 100 by providing user input, for example via the keyboard 122, mouse 124 and / or microphone 126. The system 100 may display a cursor (e.g., a mouse pointer) at a location of a point of attention of the user on the displayed user interface. In the system of Figure 1, the user can indicate the location of the point of attention using the mouse 124. Thus, the system 100 monitors the location of the point of attention and displays the cursor accordingly. In other examples, for instance where the user indicates the location of the point of attention using a touch input on a touchscreen display, the point of attention of the user may be monitored without necessarily displaying a cursor on the displayed user interface.

[0041] The operating system 110 may provide for conventional modes of user interaction with the system 100 via the displayed user interface using the mouse 122 and keyboard 124. However, as described below, the system 100 also provides an enhanced mode for interacting with the system 100 which leverages the functionality provided by a generative model 130. This mode may be referred to herein as an “enhanced interaction mode”. It will be understood that while the enhanced interaction mode may in some examples be provided as an option, in other examples the system may operate in the enhanced interaction mode only.

[0042] The system 100 may include or communicate with a speech recognition system 140. The speech recognition system 140 is shown in dashed lines in Figure 1 to highlight that it may either be a subsystem of the system 100 or alternatively may be provided in a remote system that the system 100 interacts with. The speech recognition system 140 may comprise a transcription subsystem 142 and / or a keyword detection subsystem 144. Speech processing systems and techniques for performing transcription and keyword detection are well known per se to those skilled in the art and will not be described in detail here.

[0043] The transcription subsystem 142 is configured to process audio data captured using the microphone 126 and to generate a corresponding transcription, e.g. in the form of text. In various example implementations, the transcription subsystem 142 may be used to transcribe voice input that conveys a user instruction with reference to the displayed user interface.Alternatively, a user instruction may be received in another way, for example as a text input via keyboard 122.

[0044] The keyword detection subsystem 144 is configured to process audio data captured by the microphone 126 to detect certain predefined keywords spoken by the user (e.g., “this”, “that”, “here”). In various example implementations, the detection of a keyword acts as a trigger event to identify that the user’s point of attention, as indicated by the cursor, is located at a part of the displayed user interface that the user wishes to refer to.

[0045] In some examples, the transcription subsystem 142 and keyword detection subsystem 144 may run at the same time, for example in different threads. That is, the system may generate a transcription based on received audio data while at the same time monitoring the received audio data for keywords.

[0046] The system 100 also includes a prompt generation system 150 configured to form a prompt for the generative model 130, based on user input received by the system 100. The prompt may comprise instructions to the generative model to generate computer code (“code”) configured to perform the user’s instruction.

[0047] As shown, the prompt generation system 150 includes an element-identification subsystem 152 configured to identify one or more elements of the displayed user interface based on a selected one or more locations of the point of attention. To identify the selected one or more locations on the displayed user interface, the system 100 may temporally correlate the location of the point of attention with one or more trigger events. In some implementations, a trigger event may comprise the detection of a keyword (e g., “this” or “that”) by the keyword detection system 144. Thus, a location may be selected by the user by speaking a keyword when the point of attention, as indicated by the cursor, is located at the desired location. A location could alternatively or in addition be selected in other ways, for example by clicking a button of the mouse according to a certain pattern (e.g., double click, triple click) when the cursor is at the location that the user wishes to select.

[0048] The prompt may refer to or include the one or more elements identified by the element-identification subsystem 142. For example, the prompt generation system 150 may augment a transcription of a voice command to include or refer to the one or more elements. The prompt may refer to an element by including respective structured data for the element.

[0049] Once formed, the prompt may be sent to the generative model 130 via a communication network such as the internet. In the example of Figure 1, the generative model is hosted on a remote server 153 but it may alternatively be hosted locally at the system 100. Theprompt may be sent via an application programming interface (API) request 154, and the resulting code may be received via an API response 156. The prompt may include instructions to generate the code in any suitable language, e.g. a scripting language such as JavaScript, Python or Perl.

[0050] In some examples, the code that is generated by the generative model may be validated and / or altered by a dynamic code analyser before it is executed, either locally by the system 100 and / or remotely. Suitable techniques for validating and / or altering code to check for and reduced / avoid errors are known per se to those skilled in the art and will not be described here.

[0051] The system 100 includes an appropriate runtime environment 155 configured to execute the finalised code. In various implementations, the runtime environment may for example comprise a Javascript engine such as node.js, a Python interpreter, NET Common Language Runtime etc.

[0052] The generative model may be a trained large language model which is suitable for generating code based on user instructions, e.g. an appropriate auto-regressive transformer-based neural network model. Suitable models are known per se and will not be described in detail here. Reference is directed to: A. Vaswani et al., Attention is all you need, in Advances in neural information processing systems, pages 5998-6008, 2017; J. Hoffmann et al., Training compute-optimal large language models. arXiv preprint arXiv:2203.15556, 2022; Colin Raffel et al., Exploring the limits of transfer learning with a unified text-to-text transformer. arXiv preprint arXiv: 1910.10683, 2019; Daniel Adiwardana et al., Towards a human-like open-domain chatbot. CoRR, abs / 2001.09977, 2020; Tom B Brown et al., Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020; J. Austin et al., Program synthesis with large language models. arXiv preprint arXiv:2108.07732, 2021; M. Chen et al., Evaluating large language models trained on code. arXiv preprint arXiv:2107.03374, 2021, and Yujia Li et al., Competition-level code generation with AlphaCode. Science 378, 1092-1097 (2022). In some implementations described herein, the generative model may be a multimodal large language model.

[0053] As discussed above, the system 100 provides for an enhanced mode of interacting with the system 100 which leverages the functionality of the generative model 130. The system 100 may provide the option to enter the enhanced interaction mode via a special hardware buttonor key, which may be a key on the keyboard or a button on the mouse. In some examples, the enhanced interaction mode may be entered by pressing a particular key or key combination (e.g. Ctrl -tab), or a combination of key and mouse button presses, or by selecting a button, user interface element, or other virtual selector provided in a user interface (e.g., displayed upon a touch-sensitive display screen). A sound or visual indication may be provided to indicate that the enhanced interaction mode has been activated. For example, the cursor may change appearance, e.g. it may change colour.

[0054] In some examples, the system 100 may include or interface with a structured data generation system 160. The structured data generation system 160 may be configured to generate structured data for particular resources stored on and / or displayed by the system 100, e.g. for resources that are not already represented with structured data. A resource may be any data that is accessible to the system 100, e.g. a file stored on the system 100 or an element displayed on the displayed user interface.

[0055] The structured data generation system 160 may be configured to perform optical character recognition and / or layout analysis on a resource to generate respective structured data. The structured data for a resource may comprise one or more structured data elements. For example, the structured data may comprise markup language content comprising one or more markup language elements. In some examples, the structured data may comprise HTML content comprising one or more HTML elements.

[0056] In some examples, the structured data generation system 160 may comprise a visual text processing system configured to process image data displayed on the displayed user interface in order to provide structured data (e.g. HTML content) for elements of the displayed user interface. The structured data may define a region (e.g. bounding box) containing the displayed element. The structured data may include a reference (e g. identification label) for the displayed element. The structured data may include text content obtained by performing optical character recognition on the displayed element.

[0057] The structured data generation system 160 is shown in dashed lines in Figure 1 to highlight that it may either be a subsystem of the system 100 or alternatively may be provided in a remote system that the system 100 interacts with.

[0058] Figure 2(a) and 2(b) show a first example scenario illustrating use of the enhanced interaction mode of the system 100. Figure 2(a) shows a region of a displayed user interfaceincluding display content and cursor 220. As shown, in this example the display content includes a first element in the form of a first button 212, and a second element in the form of a second button 214. The elements of the display content may be referenced by the system 100 in any suitable way, e.g. using a markup language such as Hypertext Markup Language (HTML), Extensible Markup Language (XML) or Extensible Application Markup Language (XAML), within object hierarchies, or as pixel data. The display content may be provided in any suitable context, e.g. within a browser or other application running in a desktop environment. In some examples, the display content 212, 214 may be shown within a tab of a web browser application.

[0059] The user initiates enhanced interaction mode and provides an instruction with reference to the displayed user interface. The instruction may comprise a voice command. For example, the user may point to the first button 212 using the cursor 220 and say “make this button’s background blue please”.

[0060] The word “this” is a predefined keyword, which is recognised by the keyword detection subsystem 144. The element-identification subsystem 152 identifies the location of the cursor 220 when the keyword “this” is spoken. Since this location is within the region of the displayed user interface defined by the first button 212, the element-identification subsystem 152 identifies the first button 212 as the element corresponding to “this”. For instance, in the case that the first button 212 is defined using HTML, the element-identification subsystem 152 may identify the HTML element for the first button as a reference to the selected element of the displayed user interface. The HTML element for the first button may, for example, be “<button id- ’defaultButtonE”>< / button>”.

[0061] Visual highlighting may be used to assist the user in operating the enhanced interaction mode. For example, upon identifying that the user wishes to refer to the first button, the system 100 may visually highlight that element, e.g. as shown in Figure 2(b).

[0062] The prompt generation system 150 forms a prompt for the generative model 130 based on the input received from the user. The prompt may include an initial template portion (e.g. “Generate JavaScript code to ”), together with an augmented transcription of the user’s voice command. The transcription is generated by the transcription subsystem 142 and is augmented by the prompt generation system 150 to include the reference to the first button 212 that was identified by the element-identification subsystem.

[0063] In some examples, the transcription may be augmented “on the fly” as the transcription is generated. For example, once the enhanced interaction mode has been initiated, the prompt generation system 150 may cause the transcription subsystem 142 to begin listening for a user voice command. At the same time, prompt generation system 150 may cause the keyword detection system to begin listening for keywords. Transcription may begin as soon as the user voice command is received. When the keyword “this” is detected, the existing transcription may be immediately augmented to include a reference to the first button 212 (e.g. the HTML element for the first button) at the current point in the transcription. The transcription then continues based on the remainder of the user’s voice command.

[0064] In other examples, the transcription may be generated in full (e.g., “make this button’s background blue please”) and then subsequently augmented based on the elements identified by the element-identification subsystem 152. In this case, the prompt generation system 150 may include a reference to the first button 212 (e.g. the HTML element for the first button) at a location in the transcription corresponding to the detected keyword “this”, e.g. immediately after the keyword appears in the transcription.

[0065] The augmented transcription may thus comprise the following text: “make this <button id=”defaultButtonE”>< / button> button’s background blue please”. The final prompt may comprise the initial template portion (e.g. “Generate JavaScript code”) together with the augmented transcription.

[0066] The word “please” at the end of the voice command is an example of an execution keyword which may be detected by the keyword detection subsystem 144. In response to recognising the execution keyword, the system 100 sends the finalised prompt to the generative model 130 via an API request, and receives the requested code in an API response. In this example, the API response may comprise the following JavaScript code:

[0067] / / Access the button element by its ID

[0068] const buttonE = document.getElementBylD defaultButtonE’);

[0069] / / Set the background color of the button to blue

[0070] buttonE. style. backgroundColor=’blue’;

[0071] Once the API response is received, the system 100 may execute the code using the runtime environment 155, thereby changing the background color of the first button to blue.

[0072] In some examples, the prompt may further include an instruction to include one or more commands in the code that is generated by the generative model 130 so as to generate a log explaining the action taken. In the present example, the resulting log may for example comprise “I have changed the background color of the selected button to blue”. In some implementations, the system 100 may be configured to output the log as a synthesized voice output and / or in text form on the displayed user interface.

[0073] In another example, the user may point to the second button 214 using the cursor 220 and say, “make this button’s text red please”. As in the previous example, detection of the keyword “this” acts as a trigger event. The system 100 identifies the location of the cursor 220 at the time of the trigger event and thus identifies that the user is referring to the second button 214. The prompt generation system 150 forms a prompt for the generative model 130 including the initial template portion (“Generate JavaScript code to “) and an augmented transcription of the user’s voice command. The transcription may be augmented “on the fly” by the prompt generation system 150 to include a reference to the second button 214 (e.g. the HTML element for the second button 214). In response to recognising the execution keyword “please”, the system 100 sends the prompt to the generative model 130 and receives the generated code in an API response. The system 100 executes the code using the runtime environment 155, thus making the button’s text red.

[0074] More generally, any attribute of any element displayed on the user interface may be altered using the enhanced mode by way of an appropriate voice command. For example, where the element is referenced by a structured data element, the structured data element may be included in the prompt that is sent to the generative model 130. In this way, the code that is generated by the generative model 130, when executed by the system 100, can change one or more attributes of the structured data element by referencing that element.

[0075] Transferring data between resources

[0076] Figure 3(a)-3(d) show a second example scenario illustrating use of the enhanced interaction mode of the system 100. As shown, the displayed user interface shows an image 301 comprising image regions showing “British Museum” 302, “6 Pancras Square” 304, “Latest Movies” 306, and “Arthur Bums” 307 respectively. In this example, the image 301 is an image of handwritten content on a whiteboard. For example, a human may have hand-written this texton a whiteboard (or chalkboard or similar) and the image 301 can be a photograph or other image that includes or depicts the whiteboard.

[0077] The structured data generation system 160 may perform optical character recognition (OCR) and layout analysis on the image 301 to determine structured data (e.g. markup language content such as HTML content) for the image regions 302, 304, 306, 307.

[0078] In particular, the structured data generation system 160 may generate markup language elements for the image regions 302, 304, 306, 307 including respective identification labels, positional information for bounding boxes for each, and OCR-generated text. For example, the system 100 may determine an HTML element for “British Museum” as follows:

[0079] <div id=”OCRl” style=”left: 289px; top: 198.66px; width: 425px; height:50px”>British Museum

[0080] The generated HTML elements thus act as references to the respective elements of the displayed user interface for “British Museum” 302, “6 Pancras Square” 304, “Latest movies” 304 and “Arthur Bums” 307 respectively. Additionally, in this example, each of the generated HTML elements include corresponding OCR-generated text.

[0081] The user initiates the enhanced interaction mode and provides a voice command with reference to the displayed user interface. The voice command may be “show me, using Google Maps, how to go from here to there by London underground please”. As the user speaks the command, the user may use the cursor 308 to point to “British Museum” when saying the word “here” and to “6 Pancras Square” when saying the word “there”.

[0082] As in the previous example, the system 100 temporally correlates the location of the user’s point of attention as indicated by the cursor 308 with the detection of predefined keywords in the voice command by the keyword detection subsystem 144.

[0083] In this case, the words “here” and “there” are predefined keywords. The keyword detection subsystem 144 recognises that the word “here” is a keyword. The elementidentification subsystem 152 identifies the location of the cursor 308 at the time that the keyword “here” is spoken. Since this location is within the bounding box defined by the HTML element for “British Museum”, the element-identification subsystem 152 determines that the user is referring to the element of the displayed user interface corresponding to the HTML element for “British Museum”.

[0084] The keyword detection subsystem 144 further recognises that the word “there” is a keyword. The element-identification subsystem 152 identifies the location of the cursor 308 at the time that the keyword “there” is spoken. Since this location is within the bounding box defined for the HTML element for “6 Pancras Square”, the element -identification subsystem 1 2 determines that “there” refers to an element of the displayed user interface corresponding to the HTML element for “6 Pancras Square”.

[0085] Visual highlighting may be used to assist the user in operating the enhanced interaction mode. For example, upon identifying that the user wishes to refer to “British Museum”, the system 100 may visually highlight that element, e.g. as shown in Figure 3(b). Upon subsequently identifying that the user also wishes to refer to “6 Pancras Square”, the system 100 may visually highlight that element, e.g. as shown in Figure 3(c). Different visual highlighting may be used to distinguish different elements.

[0086] The prompt generation system 150 forms a prompt for the generative model 130 based on the input received from the user. The prompt may include an initial template portion (e.g. “Generate JavaScript code to ”), together with an augmented transcription of the user’s voice command. In forming the augmented transcription, the prompt generation system 150 may augment a transcription generated by the transcription subsystem 145 to include the HTML element for “British Museum” when the user says the keyword “here”, and to include the HTML element for “6 Pancras Square” when the user says the keyword “there”.

[0087] Note that in this example, the HTML elements for “British Museum” and “6 Pancras Square” include the respective text determined by the OCR process. That is, in this example, each HTML element not only refers to a particular element of the displayed user interface, but also defines content relating to that element, which is included in the prompt. Generally, by providing relevant content in data structures (e.g. HTML elements) corresponding to elements of the displayed user interface, the system 100 ensures that the generative model 130 is provided with the information that it needs to perform various requested instructions by the user. In this case, since HTML elements that include the text obtained by the OCR process are included in the prompt, the generative model 130 is provided with that text for use in generating code to perform the instruction requested by the user.

[0088] As in the first example scenario, the word “please” may be an execution keyword which may be detected by the keyword detection subsystem 144. In response to recognising theexecution keyword in the voice command, the system 100 sends the finalised prompt to the generative model 130 via an API request, which generates code (e.g. JavaScript code) to obtain directions, with Google Maps, from the British Museum to 6 Pancras Square using London Underground. Once the API response containing the code is received, the system 100 executes the code using the runtime environment 155, thereby launching Google Maps and providing a request to that application to show the requested directions. The result is shown in Figure 3(d). In browser-based implementations, such as that shown in Figure 3(a), the prompt may include an instruction to open the requested application (e.g. Google Maps) in a different tab to the tab in which the image 301 was shown.

[0089] In another example, the voice command may be “search for this on YouTube”. As the user says the word “this”, the user may use the mouse cursor 308 to point to “Latest Movies” 306.

[0090] The words “this” is an example of a predefined keyword, which is recognised by the keyword detection subsystem 144. The element -identification subsystem 152 identifies the location of the cursor 308 when the keyword “this” is spoken. Since this location is within the bounding box defined for the element for “Latest Movies”, the element-identification subsystem 152 determines that “this” refers to an element of the displayed user interface corresponding to the HTML element for “Latest Movies” 306.

[0091] Upon identifying that the user wishes to refer to “Latest Movies”, the system 100 may visually highlight that element in a similar fashion to that shown in Figure 3(b) and Figure 3(c).

[0092] The prompt generation system 150 forms a prompt for the generative model 130 based on the input received from the user. The prompt may include an initial template portion (e.g. “Generate JavaScript code to ”), together with an augmented transcription of the user’s voice command. The transcription may be generated by the transcription subsystem 142 and augmented by the prompt generation system 150 as described above to include the HTML element for “Latest Movies”.

[0093] In response to recognising the execution keyword “please”, the system 100 sends the finalised prompt to the generative model 130 via an API request, and receives the requested code in an API response. Once the API response is received, the system 100 executes the code using the runtime environment 155, thereby executing a search on YouTube for “Latest Movies”.

[0094] In some examples, more complex searches may be instructed. For instance, the voice command may comprise an instruction to search a collaborative document processing system for files that were developed together with a particular individual. The voice command may be “search Google Docs for documents that I worked on with this person please”. When saying the word “this”, the user may use the mouse to identify another user of the collaborative document processing system, e.g. by pointing the mouse to the name “Arthur Bums” 307 in the image 301. The prompt generation system 150 forms a prompt by augmenting a transcription of the voice command using the HTML element for “Arthur Burns” as described above. The prompt may be sent to the generative model 130, which generates code that is then executed by the system 100 to perform the requested search.

[0095] The examples of Figures 3(a)-3(d) illustrate how the enhanced interaction mode may be used to transfer data from a particular resource (in this case the image 301) to another resource, e.g. to a web application such as Google Maps, YouTube or Google Docs.

[0096] Figure 4(a)-4(f) show a third example scenario illustrating use of the enhanced interaction mode of the system 100. This scenario provides a further example of how the enhanced interaction mode may be used to transfer data from a first resource of the system 100 to a second resource. In particular, this scenario provides an illustrative example of how the enhanced interaction mode of the system may be used for assisted form filling, or more generally, to complete input fields based on information from another resource.

[0097] The first resource may be an image file containing an image 400 as shown in Figure 4(a). The image may for example represent a photograph of handwritten text on a piece of paper. For example, the image 400 can be a photograph or a scan of an adhesive sticky note that contains handwritten textual content. The second resource may be a web browser, which in the example of Figure 4(a) is displaying a webpage which allows a user to complete a form 420.

[0098] Figure 4(b) shows a folder of the filesystem of the system 100 displaying a thumbnail for the image 400. The structured data generation system 160 may perform optical character recognition (OCR) on the image file 410 to determine structured data (e.g. an HTML element) for the text shown in the image 400.

[0099] As shown in Figure 4(a), the form 420 has various input fields (e.g. “Reference Number Client”, “Name”, “Date of Birth” etc). The form may be represented by a structured data element (e.g. an HTML <form> element>, which may include various sub-elements for eachinput fields. The HTML element for the form may define a region (e g. a bounding box) within which the form is located.

[0100] The user initiates the enhanced interaction mode and provides a voice command with reference to the displayed user interface. The voice command may be “fill out this form with that information, please”. As the user speaks the command, the user may use the cursor 424 to point to the form 420 when saying the word “that”, and to the image file 410 when saying the word “this”.

[0101] The words “this” and “that” are predefined keywords. The keyword detection subsystem 144 recognises that the word “this” is a keyword. The element-identification subsystem 152 identifies the location of the cursor 424 when the keyword “this” is spoken. Since this location corresponds to the region (e.g. bounding box) defined for the form, the elementidentification subsystem determines that “this” refers to an element of the displayed user interface corresponding to the structured data element (e.g. HTML element) for the form 420, which as discussed above also includes various sub-elements defining the various input fields of the form 420.

[0102] The element-identification subsystem 152 also identifies the location of the cursor 424 when the keyword “that” is spoken. Since the location corresponds to the image file 410, the element-identification subsystem identifies that “that” corresponds to image file 410.

[0103] As in the previous examples, visual highlighting may be used to assist the user in operating the enhanced interaction mode. For example, upon identifying that the user wishes to refer to the form 420, the system 100 may visually highlight that element, e.g. as shown in Figure 4(c). Upon identifying that the user also wishes to refer to the file 410, the system 100 may visually highlight that element, e.g. as shown in Figure 4(d). Different visual highlighting may be used to distinguish different elements.

[0104] The prompt generation system 150 forms a prompt for the generative model 130 based on the input received from the user. The prompt may include an initial template portion (e.g. “Generate JavaScript code to ”), together with an augmented transcription of the user’s voice command.

[0105] In particular, since the element-identification subsystem 152 identified that “that” corresponds to the image file 410, the prompt generation system 150 augments the transcriptionbased on the structured data for the image 400, which as discussed above includes the text determined by performing OCR on the image 400.

[0106] Since the element-identification system 152 identified that “this” corresponds to the form 420, the prompt generation system 150 also augments the transcription based on the HTML content for the form, which as discussed above includes sub-elements for the input fields of the form 420.

[0107] As discussed above, the transcription may be augmented “on the fly” as the transcription is generated by the transcription subsystem 142, i.e. the transcription may be augmented to include HTML content as the respective keywords are spoken by the user.Alternatively the transcription may be first generated in full and then later augmented at locations corresponding to the positions of the keywords in the transcription.

[0108] In either case, the prompt includes the text determined by performing OCR on the image 400, as well as all of the sub-elements for the input fields of the form. Thus, the generative model 130 can generate code to populate the relevant input fields of the form based on the information shown in the image 400.

[0109] In response to recognising the execution keyword “please” in the voice command, the system 100 sends the finalised prompt to the generative model 130 via an API request, and receives the requested code in an API response. Once the API response is received, the system 100 executes the code, thereby filling out the relevant fields of the form as shown in Figure 4(e).

[0110] In some examples, the user may have a continued interaction with the system 100 using the enhanced interaction mode. As shown in Figure 4(e), one field (“address”) has not been filled out, as this information was not present in the text included in the image 400. The user may provide the voice command “add the address here, it is London UK nlc ABC, please”. The user may use the cursor 424 to point to the form when saying the keyword “here”. The prompt generation system 150 forms a prompt by augmenting a transcription of the voice command to include the HTML content for the form 420 at the relevant location of the transcription.

[0111] Thus, the prompt contains the information which the generative model 130 needs to generate code to fill out the address field of the form. In response to recognising the execution keyword “please” in the voice command, the system 100 sends the finalised prompt to the generative model 130 via another API request, and receives the requested code in acorresponding API response. Once the API response is received, the system 100 executes the code, thereby fdling out the address field of the form as shown in Figure 4(f).

[0112] As discussed above, the system 100 may provide visual, audio and / or text indications to help guide the continued interaction between the user and the system 100. For example, upon identifying that the user wishes to refer to an element (e.g. the form), visual highlighting may be used to highlight it. Moreover, the prompt may include an instruction to include one or more commands in the code to generate a log explaining the action taken, which may be output as a synthesized voice output and / or in text form on the displayed user interface.

[0113] In some examples, the enhanced interaction mode may be used to fill out a form based only on information provided in the user instructions, rather than based on information in a file. Figure 5(a) shows an example of a completed form based on the voice command “help me fill out this form. The overall food quality was above average, the food freshness was good, the menu was super diverse, the service was pretty poor though, and the people were not polite at all”. As shown in Figure 5(b), a visual indication (e.g. a line joining the individual form items, which may be animated) may also be provided to indicate to the user that the form has been completed.

[0114] Figure 6(a) and 6(b) show a further example scenario illustrating use of the enhanced interaction mode of the system 100. In Figure 6(a), the displayed user interface shows a music player 610. The music player is an example of an element of the displayed user interface whose appearance is determined by a first set of attributes. For example, the appearance of the music player 610 may be defined by structured data, e.g. HTML or XML content, including a set of attributes which define e.g. the text and background colors, font, font size, as well as the location and size of the music player 610 and its sub-elements on the displayed user interface.

[0115] Figure 6(a) also shows the contents 630 of a text file 620 including accessibility guidelines for use in styling the music player 610. The text file 620 is an example of an element of the displayed user interface which defines a second set of attributes which are different to the first set of attributes. For example, the accessibility guidelines may comprise a further set of attributes for particular text and background colors, font, font size etc, which are different to those currently used in the music player 610. The text in the text file 620 may be represented by structured data, or structured data may be generated for it using the structured data generation system 160.

[0116] As discussed below, the enhanced interaction mode may be used to alter the first set of attributes based on the second set of attributes, thereby updating the appearance of the element of the displayed user interface defined by the first set of attributes.

[0117] For example, the user may provide a voice command to “style this music player with those accessibility guidelines please”. The word “this” is a keyword, which is recognised by the keyword detection subsystem 144. The element -identification subsystem 152 identifies the location of the cursor 640 when the keyword “this” is spoken. Since this location overlaps with the displayed music player, the element-identification subsystem determines that “this” refers to the displayed music player.

[0118] The element-identification subsystem 152 also recognises the keyword “those”. The element-identification subsystem 152 identifies the location of the cursor 640 when the keyword “those” is spoken. Since this location overlaps with the text file 620, the elementidentification subsystem identifies that “those” refers to the text file 620.

[0119] The prompt generation system 150 forms a prompt for the generative model 130 to generate code to perform the user’s instruction. An augmented transcription is generated and included in the prompt. In particular, the prompt generation system 150 may include the structured data for the music player 610 in the prompt at a location of the transcription corresponding to the detected keyword “this”. The prompt generation system 150 also includes the structured data representing the accessibility guidelines at a location of the transcription corresponding to the detected keyword “those”. In this way, the prompt includes the information which the generative model 130 needs to generate code to alter the set of attributes defined in the structured data for the music player 610 based on the set of attributes defined in the accessibility guidelines 620.

[0120] In response to recognising the execution keyword “please” in the voice command, the system 100 sends the finalised prompt to the generative model 130 via an API request, and receives the requested code in an API response. Once the API response is received, the system 100 executes the code, thereby adjusting the attributes of the music player in accordance with the accessibility guidelines 620. The result is shown in Figure 6(b).

[0121] Editing images

[0122] Figures 7(a)-(c) illustrate a further example scenario illustrating use of the enhanced interaction mode of the system 100 for image editing. Figure 7(a) shows an image 702,which may for example be shown in a tab of a web browser application running on the system 100. The image is defined using structured data (e.g. HTML content), which may include a structured data element (e.g. HTML element) for a sub-image 704 of the image 702. The structured data element may define an ID for the sub-image 704, as well as one or more attributes defining the size of the sub-image (e.g. its width and height).

[0123] The user initiates the enhanced interaction mode and provides a voice command with reference to the displayed user interface. The voice command may for example be “make this elephant bigger please”. As the user speaks the command, the user may use the cursor to point to the sub-image 704 when saying the word “this”.

[0124] The word “this” is an example of a predefined keyword which is recognised by the element-identification subsystem 152. The element-identification subsystem 152 identifies the location of the cursor when the keyword “this” is spoken. Since this location corresponds to the sub-image 704, the element-identification subsystem 152 identifies that “this” corresponds to the sub-image 704. As in previous examples, upon identifying that the user wishes to refer to the sub-image 704, the system 100 may visually highlight that element.

[0125] The prompt generation system 150 forms a prompt for the generative model 130 to generate code to perform the user’s instruction. The prompt generation system 150 may augment a transcription of the user’s voice command that is generated by the transcription subsystem 142. In particular, the prompt generation system 150 may augment the transcription based on the structured data element for the sub-image 704. For example, the prompt generation system 150 may insert the structured data element for the sub-image 704 at a location of the transcription corresponding to the detected keyword “this”. In this way the prompt includes a reference to the sub-image 704, as well as its attributes, at a relevant location in the prompt.

[0126] In response to recognising the execution keyword “please” in the voice command, the system 100 sends the finalised prompt to the generative model 130 via an API request. The generative model 130 generates code based on the prompt. For example, the generated code may access the sub-image 704 via the ID included in its HTML element, and may increase the width and height attributes of the sub-image 704, for instance by 20%. The system 100 receives the requested code in an API response. Once the API response is received, the system 100 executes the code, thereby adjusting the size of the sub-image 704. The result is shown in Figure 7(b).

[0127] In some examples, the prompt generation system 150 may be configured to provide an instruction in the prompt to generate code to provide a visual control (e.g. a slider or colour picker) to permit further control of the attribute that the user wishes to adjust. Continuing the present example, the user may use the visual control to further enlarge the sub-image 704 to the size shown in Figure 7(c).

[0128] Other attributes of the sub-image 704 and / or attributes of other elements of the image 702 may be adjusted in a similar way. For example, the user instruction may be to make another elements of the image 702 (e.g. mountains 708, which may be defined by another structured data element) smaller, or to move one element of the image to the front or behind other elements. In another example, the user instruction may be to make a hidden element visible. Figure 7(d) shows an example in which a hidden element (fireworks 710) has been made visible.

[0129] Editing Charts

[0130] The enhanced interaction mode of the system 100 may be used to edit graphs / charts in a similar way to the way that it is used to edit images as described above with reference to Figures 7(a)-(c). In particular, the system 100 may be used to adjust various attributes of a chart based on a user command. Figure 8(a) shows an example of a chart showing temperature in degrees Celsius. The user moves the cursor over the chart and provides the command “update this chart to use degrees Fahrenheit”. The system 100 identifies structured data for the chart and forms a prompt comprising an augmented transcription of the user’s voice command including the structured data. The generative model 130 receives the prompt and generates code for executing the user’s voice command. An example result is shown in Figure 8(b). Alternatively, the user may provide the command “please update this chart to be a line chart instead of a bar chart”. An example result for this command is shown in Figure 8(c).

[0131] Voice Programming

[0132] In some implementations, the enhanced interaction mode of the system may be used to conditionally link displayed objects, and more generally, for programming and software development.

[0133] Figure 9 illustrates an example in which the enhanced interaction mode is used to assign an action to an element 902 shown on the displayed user interface. For instance, the usermay provide the voice command “whenever T click on this button, the background color of that text box should change to any random colour”.

[0134] The user positions the cursor 906 over the button 902 when saying the word “this” and over the text box 908 when saying the word “that”. Thus, the element-identification subsystem identifies the button 902 and the text box 908 as relating to “this” and “that” respectively. The system 100 may visually highlight the button 902 and text box 908 to confirm to the user that they have been identified.

[0135] The prompt generation system 150 forms a prompt for the generative model 130 to generate code based on the user’s instruction. An augmented transcription of the user’s instruction is generated and included in the prompt. For example, the prompt generation system 150 may include a reference to the button 902 at a location of the transcription corresponding to the detected keyword “this”, and a reference to the text box 908 at a location of the transcription corresponding to the location of the keyword “that”. The reference to the button 902 may comprise an HTML element defining the button 902, and the reference to the text box 908 may comprise an HTML element defining the text box 908. Thus, the generated prompt may comprise “whenever I click this button <button id=”defaultButtonToConnect”> BUTTON < / button> the background of that text box <textarea id=”Topic”>< / textarea> should change to a random colour, please”. In some examples, the prompt generation system 150 may also include additional instructions in the prompt, for example an instruction to include code to generate a log.

[0136] In response to recognising the execution keyword “please” in the voice command, the system 100 sends the prompt to the generative model 130 via an API request. The generative model 130 may, for example, generate the following lavaScript code based on the prompt:

[0137] / / Get a reference to the button and the text area using their IDs.

[0138] const button=

[0139] document.getElementbylD defaultButtonToConnect’);

[0140] const textArea=

[0141] document.getElementbylDf Topic’ );

[0142] / / Add an event listener to the button.

[0143] / / When the button is clicked, the function inside will be executed

[0144] button. addEventListener(‘click’, () => {

[0145] / / Generate random color.

[0146] const randomColor=’#’ +

[0147] Math.floor(Math.random()*16777215).toString(16

[0148] );

[0149] / / Set the background color of the text area to the random color.

[0150] textArea.style.backgroundColor=randomColor;

[0151] appendToLog(‘javascriptLogl’, ‘The background color of the text area has been changed.’);

[0152] });

[0153] In this example, the generative model 130 has generated code which creates a “listener” by attaching an event handler to the button 902, so that each time the button 902 is clicked, the text box 908 changes to a random color.

[0154] The system 100 receives the code in an API response and executes it, thereby implementing the listener on the system 100. In some examples, a visual indication (e.g. a line joining the button to the textbox) may also be provided to indicate to the user that the requested action has been completed.

[0155] In some examples, more complex programs may be created. For instance, using the techniques described above, the user may control the system 100 to create a listener which, when the element 902 is selected, processes the text included in the textbox 908 and generates, in a further textbox 910, output conditional on that text. For instance, the user may control the system 100 to create a listener which, when the button 902 is pressed, reads the text included in the textbox 908 and generates a poem about that topic in a further textbox 910. Figure 10(a) shows an example in which the user has included the word “cat” in the textbox 908 and pressed the button 902. If the user changes the text in the textbox 908 and presses the button 902 again, another output may be generated in the further textbox 910. An example is shown in Figure 10(b).

[0156] Figures 11(a) and 11(b) illustrate another example of the use of the enhanced interaction mode of the system 100 in software development. In this case, the enhanced interaction mode of the system 100 is used to augment existing code with comments and / or log functionality. Figure 11(a) shows an example in which the displayed user interface shows code 1102 which has been written by the user. The text of the code may be represented using astructured data element (e.g. HTML element such as a or <textarea> element) which includes an identifier that acts as a reference to the code. When the user moves the cursor 1104 over the text and says, “add logs and comments to this function”, the element-identification subsystem 152 determines that the user is referring to the code 1102 using the techniques described above.

[0157] The prompt generation system 150 forms a prompt based on an augmented transcription of the voice command provided by the user. The prompt generation system includes the structured data element for the code at a location in the transcription corresponding to the position of the detected keyword “this”. The system 100 sends the finalised prompt to the generative model 130, which generates an output comprising code to replace the text contained in the structured data element with text in which logs and comments have been added to the code for the function. An example result is shown in Figure 11(b).

[0158] Document editing

[0159] Figures 12(a)-(c) illustrate examples of use of the enhanced interaction mode of the system 100 for document editing.

[0160] In the example of Figure 12(a), the displayed user interface shows different paragraphs of text 1202, 1204, 1206, 1208, which are represented by respective structured data elements (e.g. <textarea> elements) which each include a respective identifier that acts as a reference to the respective text.

[0161] The user may manipulate the text by providing voice commands which refer to paragraphs of the text using certain predefined keywords (e.g. “this” / “that”). As described above, the system 100 temporally correlates the location of the user’s point of attention as indicated by the cursor with the detection of the keywords.

[0162] For example, the user may move the cursor over the first paragraph and say “delete this paragraph”, and then move the cursor over the second paragraph and say “and also delete that one”. The system 100 identifies that “this” and “that” refer to the structured data elements for the text 1202 and 1206 respectively.

[0163] The prompt generation system 152 forms a prompt by augmenting a transcription of the user’s voice command to include the structured data element for the respective text elements, at corresponding locations in the transcription. Thus, the generative model 130 is provided with references to the structured data elements for the text to be edited.

[0164] In response to receiving the prompt, the generative model 130 generates code to perform the requested document editing command with reference to the structured data elements provided in the prompt. The result is shown in Figure 12(b), in which the text included in the structured data elements for the text 1202 and 1206 has been deleted.

[0165] Other examples may leverage the ability of generative models to alter the style of given text or to translate it from one language to another. For example, the user may move the cursor over the fourth paragraph and say “make this paragraph more formal and in French”. The prompt generation system 152 forms a prompt accordingly, which is sent to the generative model 130. The system 100 receives the code generated by the generative model 130 and executes the code, thereby performing the requested document-editing command by editing or replacing the text included in the relevant HTML element. An example result is shown in Figure 12(c).

[0166] More generally, the enhanced interaction mode may be used to edit any text shown on the displayed user interface in accordance with a user’s command. Alternatively, or in addition, the enhanced interaction mode may be used to alter the formatting of text, e.g. the font size or font type, for example by altering one or more attributes of the structured data that is passed to the generative model via the prompt.

[0167] Editing Spreadsheets

[0168] Figures 13(a) and 13(b) illustrate examples of use of the enhanced interaction mode of the system 100 for editing a spreadsheet.

[0169] In this and other examples, the user may make use of a multi-select mode in which the user is able to select more than one element of the displayed user interface with a single trigger event (e.g. single keyword). The user may use a particular button or key, or a particular combination of inputs, in order to use the multi-select mode. For example, the user may hold down the shift key while saying a particular keyword (e g. “those”) in order to indicate that the user wishes to select more than one element of the displayed user interface.

[0170] The system 100 may augment a transcription of a voice command based on structured data elements for the elements of the displayed user interface that are selected by the user. In the example of Figure 13(a), the user may activate the multi-select mode and say “write the days of the week in those cells”, whilst moving the cursor 1202 across all of the cells 1304-1310. The element-identification subsystem may thus identify the relevant elements of the displayed user interface to be those corresponding to the structured data elements for the cells1304-1310. The prompt generation system may thus generate a prompt comprising an augmented transcription including all of the structured data elements for 1304-1310 (e.g. one after the other) at a location of the transcription corresponding to the detected keyword “those”. In this way, the generative model 130 is provided with structured data for the areas of the spreadsheet referenced by the user, and thus can generate code to populate those cells with the requested information. An example result is shown in Figure 13(b).

[0171] In this and other examples, the enhanced interaction mode may make use of previous user commands and their corresponding responses when generating the response to a new command. For example, if the user provides a new voice command to “change Wednesday to be red, Friday to be blue, and Sunday to be purple”, the generative model 130 may recognise from the previous command and response that “Wednesday, “Friday” and “Sunday” correspond to the cells 1306, 1308, 1310 respectively. In this way, the generative model 130 has the information that is needed to perform the user’s command.

[0172] Providing Text Instructions

[0173] In some examples, the user may convey the user instruction using text rather than as a voice command. As illustrated in the example of Figure 14, the user makes a text selection 1402 within a document 1403 and provides a non-verbal trigger input (e.g. by pressing a particular key or button, or particular combination of keys and / or buttons, e.g. a triple click of a mouse button) when the cursor is over the selection. The selected text is thus identified by the element-identification system 152 for inclusion in the prompt for the generative model. The displayed user interface may include a selectable element 1404 configured to cause the prompt to be sent to the generative model when the selectable element is selected (e.g. clicked) by the user. The generative model 130 then generates code based on the prompt, and the system 100 executes the code. In this example, the system 100 may open relevant browser tabs for the relevant articles on the shopping website “shoppingwebsite”.

[0174] Hardware button or key

[0175] The system 100 may provide the option to enter the enhanced interaction mode via a special hardware control, e.g. a key on the keyboard or a button on the mouse. For example, the special hardware control may comprise an additional key on the keyboard, or an additional button on the mouse. In some examples the same hardware control may be used by the user to provide a trigger event to select elements on the displayed user interface, as an alternative to useof keywords, for example by selecting the button or key in accordance with a predefined pattern (e.g. by double clicking a button). In some implementations, the multi-select mode may be initiated using the same hardware control, e.g. by holding the button or key.

[0176] Prompt generation techniques

[0177] Figure 15 illustrates a method 1500 in accordance with various example implementations described in this specification. The method may be implemented by one or more computing devices at one or more locations.

[0178] The method 1500 comprises receiving 1502 user input. The user input may for example comprise a voice command or text input that conveys a user instruction with reference to a displayed user interface.

[0179] The method 1500 also includes monitoring 1504 a location of a point of attention of the user on the displayed user interface. In some implementations, the point of attention may be identified on the displayed user interface by a cursor (e.g. mouse pointer).

[0180] The user input is evaluated 1506 for one or more trigger events. For example, where the user input comprises a voice command, the voice command may be evaluated for one or more trigger events in the form of one or more keywords.

[0181] The method includes identifying 1508 one or more elements of the displayed user interface based on the location of the point of attention at one or more times. In some examples, the one or more times may be selected based on the one or more trigger events. For example, an element of the displayed user interface may be identified based on the location of the point of attention at the time of a trigger event (e.g. at the time of detection of a keyword). More generally, identifying the one or more elements may comprise temporally correlating the location of the point of attention with the one or more trigger events.

[0182] An element of the displayed user interface may be identified if the location of the point of attention at the one or more times is within the element. As described in various examples set out above, an element may be identified using structured data relating to the element. For example, the structured data may for example comprise position and / or extent information which defines the position and / or extent of the element on the displayed user interface.

[0183] However, in other examples, an identified element may comprise image data (e.g., pixel data) extracted from what is currently shown on the displayed user interface. For example,the identified element may comprise image data for a region of the displayed user interface that contains the location of the point of attention at a selected time. The size and / or shape of the region may be fixed (and thus the same for each identified element), or it may vary, in dependence on, for example, layout analysis that is performed on what is currently displayed on the displayed user interface.

[0184] In some examples, the method may include visually highlighting 1510 the identified element(s), thereby providing feedback / confirmation to the user.

[0185] The method includes forming 1512 a prompt for a generative model based on the user input and the one or more elements. Forming the prompt may include including data relating to the one or more elements in the prompt. For example, where the user input comprises a voice command, forming the prompt may comprise generating an augmented transcription of the voice command which includes data relating to the one or more elements. Where the user input comprises text and does not include a voice command, forming the prompt data may comprise inserting data relating to the one or more elements into the text

[0186] As described in various examples set out above, the data that is included in the prompt for an identified element may comprise structured data. The structured data may be innate to the element or it may have been generated by a structured data generation system (e.g. by OCR and / or layout analysis).

[0187] In some implementations, structured data may be generated by a structured data generation system on demand. For example, in the case that an identified element comprises image data, structured data may be generated based on the image data, e.g. using OCR and / or layout analysis. The generated structured data may then be included in the prompt.

[0188] Alternatively, however, image data (e.g. pixel data) for a region of the displayed user interface may be included in the prompt directly. In these implementations, the generative model may comprise a multimodal language model which is configured to process input text and image data. Multimodal language models are known per se and will not be described here.Examples include those described in Andreas Steiner et al, “PaliGemma 2: A Family of Versatile VLMs for Transfer”. arXiv preprint arXiv: 2412.03555 (2024), Gemini Team, Google “Gemini 1.5 Unlocking multimodal understanding across millions of tokens of context”. arXiv preprint arXiv: 2403.05530 (2024), Li, J. et al, BLIP -2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023);Albayrak, J.B. et al., Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems 35, 23716- 23736 (2022); Jia, C. et al, Scaling up visual and vision-language representation learning with noisy text supervision. In: International Conference on Machine Learning, pp. 4904-4916. PMLR (2021); OpenAI: GPT-4 technical report. arXiv preprint arXiv:2303.08774 (2023); and Lu, J. et al, Unified-IO: A unified model for vision, language, and multi-modal tasks. arXiv preprint arXiv:2206.08916 (2022).

[0189] In the case that an identified element comprises image data that is provided to the generative model, the prompt may include instructions to generate code to modify that image data in accordance with the user instruction.

[0190] Referring again to Figure 15, the method 1500 further comprises sending 1514 the prompt to a generative model so as to generate code corresponding to the user instruction, and executing 1516 the code.

[0191] The generative model may be hosted on a remote server. However, in some embodiments, the prompt may be sent to a generative model which is hosted locally. In some examples, the code that is generated by the generative model may be validated and / or altered by a dynamic code analyser before it is executed, either locally and / or remotely. Suitable techniques for validating and / or altering code to check for and reduce / avoid errors are known per se to those skilled in the art and will not be described here.

[0192] In some examples, the method may further comprise 1518 providing a visual and / or audio indication that the code has been executed. For example, the system may provide a confirmation in the form of a synthesized voice output and / or text output, which may be formed based on a log which is produced by the executed code. Alternatively, or in addition, the system may provide a visual indication such as a line joining identified elements together.

[0193] Figures 16 and 17 illustrate another example method 1700 for forming a prompt for a generative model. Referring to Figure 16, in this example, the prompt may include a number of prompt elements 1604 reflecting the user’s interaction with the system over time. Thus, the prompt may be formed based on a time-based logging approach. Prompt items may include transcriptions 1606 of part of the voice input, or structured data 1608 or text selections 1610 for elements shown on the displayed user interface. Each prompt item is timestamped and thus the prompt as a whole acts as a detailed log which the generative model may use to infer user intent and generate an appropriate response.

[0194] Figure 17 illustrates steps of the example method 1700. As shown, the method includes monitoring 1702 a location of a point of attention of a user on a displayed user interface. One or more prompt items 1704 are determined, each prompt item including a timestamp. One or more of the prompt items may be determined based on the location of the point of attention of the user at a respective time. For example, structured data (e.g. a markup language element) for an element of the displayed user interface may be obtained based on the location of the point of attention of the user at a respective time. The structured data may be included in a prompt item, which may be time-stamped based on the respective time. In addition, one or more of the prompt items may include a transcription of part of a voice input received from a user. Transcriptions may for example be timestamped based on the starting time of the transcription. The method 1700 further comprises forming 1706 a prompt for a generative model based on the one or more prompt items. The prompt may be sent to the generative model.

[0195] In some examples, the voice input may convey an instruction, and the prompt may comprise a prompt to generate computer code to perform the instruction. However, alternatively, the voice input may form part of a dialog based (e.g. chatbot) interaction with the generative model. For instance, in the example of Figure 16, the user may hover over “London Eye” whilst providing the voice input “Is this place worth visiting”, thus generating the prompt items 1604. The resulting prompt may then be sent to the generative model which generates a corresponding response.

[0196] Figure 18 illustrates another example method 1800 of forming a prompt. The method includes obtaining 1802 a context stream comprising a stream of context data over time. The context stream may for example comprise a data stream indicating cursor locations (e.g. mouse pointer positions) on a display over time. More generally, the context stream may indicate the location of a point of attention of a user over time. User input is received 1804 which refers to the context stream. For example, the user input may comprise voice input which refers to one or more positions of the cursor using certain words such as “here” or “there”. The user input is evaluated 1806 for one or more trigger events. For example, in the case that the user input comprises a voice input, the user input may be processed to detect one or more predefined keywords (e.g. “here”, “there”) in the voice input.

[0197] The method further comprises determining 1808 one or more prompt items by sampling from the context stream in temporal correspondence with the one or more triggerevents. For example, a context data item may be selected from the context stream if it coincides with a trigger event. Determining that a context data item coincides with a trigger event may be performed in real-time as context data items and trigger events are received, or it may be performed after the context data item and trigger event are received, e.g. by comparing time stamps for the context data items with time stamps for the trigger events. In the case that the user input comprises a voice input, a trigger event may comprise detection of a predefined keyword in the voice input.

[0198] The selected one or more context data items are used to determine prompt items. For example, where a selected context data item comprises a cursor position, a prompt item may be selected based on the cursor position. A prompt item may for example comprise structured data for one or more visual components (e.g. one or more user interface elements) that are displayed on a display. In this case, structured data for a visual component may be selected as a prompt item if the cursor position is within the visual component. Alternatively, a prompt item may comprise image data representing one or more image regions shown in a display. In this case, the image data may be selected as a prompt item if the cursor position is within the image region.

[0199] The method further comprises forming 1810 a prompt for a generative model using the one or more prompt items. In some examples, forming the prompt may comprise inserting prompt items into text corresponding to a user instruction conveyed by the user input. Where the user input is a voice input, the text may comprise a transcription of the voice input, which may be augmented using the prompt items.

[0200] Figure 19 depicts a flowchart of a method 1900 for training one or more machine-learned models according to aspects of the present disclosure. One or more portion(s) of example method 1900 can be implemented by a computing system that includes one or more computing devices such as, for example, computing systems described with reference to the other figures. Each respective portion of example method 1900 can be performed by any (or any combination) of one or more computing devices. Moreover, one or more portion(s) of example method 1900 can be implemented on the hardware components of the device(s) described herein, for example, to train one or more systems or models. Figure 19 depicts elements performed in a particular order for purposes of illustration and discussion. Those of ordinary skill in the art, using the disclosures provided herein, will understand that the elements of any of the methods discussedherein can be adapted, rearranged, expanded, omitted, combined, or modified in various ways without deviating from the scope of the present disclosure. Figure 19 is described with reference to elements / terms described with respect to other systems and figures for exemplary illustrated purposes and is not meant to be limiting. One or more portions of example method 1900 can be performed additionally, or alternatively, by other systems.

[0201] At 1902, example method 1900 can include obtaining a training instance. A set of training data can include a plurality of training instances divided between multiple datasets (e.g., a training dataset, a validation dataset, or testing dataset). A training instance can be labeled or unlabeled. Although referred to in example method 1900 as a “training” instance, it is to be understood that runtime inferences can form training instances when a model is trained using an evaluation of the model’s performance on that runtime instance (e.g., online training / leaming). Example data types for the training instance and various tasks associated therewith are described throughout the present disclosure.

[0202] At 1904, example method 1900 can include processing, using one or more machine-learned models, the training instance to generate an output. The output can be directly obtained from the one or more machine-learned models or can be a downstream result of a chain of processing operations that includes an output of the one or more machine-learned models.

[0203] At 1906, example method 1900 can include receiving an evaluation signal associated with the output. The evaluation signal can be obtained using a loss function. Various determinations of loss can be used, such as mean squared error, likelihood loss, cross entropy loss, hinge loss, contrastive loss, or various other loss functions. The evaluation signal can be computed using known ground-truth labels (e.g., supervised learning), predicted or estimated labels (e.g., semi- or self-supervised learning), or without labels (e.g., unsupervised learning). The evaluation signal can be a reward (e.g., for reinforcement learning). The reward can be computed using a machine-learned reward model configured to generate rewards based on output(s) received. The reward can be computed using feedback data describing human feedback on the output(s).

[0204] At 1908, example method 1900 can include updating the machine-learned model using the evaluation signal. For example, values for parameters of the machine-learned model(s) can be learned, in some embodiments, using various training or learning techniques, such as, for example, backwards propagation. For example, the evaluation signal can be b ackprop agatedfrom the output (or another source of the evaluation signal) through the machine-learned model(s) to update one or more parameters of the model(s) (e.g., based on a gradient of the evaluation signal with respect to the parameter value(s)). For example, system(s) containing one or more machine-learned models can be trained in an end-to-end manner. Gradient descent techniques can be used to iteratively update the parameters over a number of training iterations. In some implementations, performing backwards propagation of errors can include performing truncated backpropagation through time. Example method 1900 can include implementing a number of generalization techniques (e.g., weight decays, dropouts, etc.) to improve the generalization capability of the models being trained.

[0205] In some implementations, example method 1900 can be implemented for training a machine-learned model from an initialized state to a fully trained state (e.g., when the model exhibits a desired performance profile, such as based on accuracy, precision, recall, etc ).

[0206] In some implementations, example method 1900 can be implemented for particular stages of a training procedure. For instance, in some implementations, example method 1900 can be implemented for pre-training a machine-learned model. Pre-training can include, for instance, large-scale training over potentially noisy data to achieve a broad base of performance levels across a variety of tasks / data types.

[0207] In some implementations, example method 1900 can be implemented for fine-tuning a machine-learned model. Fine-tuning can include, for instance, smaller-scale training on higher-quality (e.g., labeled, curated, etc.) data. Fine-tuning can affect all or a portion of the parameters of a machine-learned model. For example, various portions of the machine-learned model can be “frozen” for certain training stages. For example, parameters associated with an embedding space can be “frozen” during fine-tuning (e.g., to retain information learned from a broader domain(s) than present in the fine-tuning dataset(s)). In some implementations, example method 1900 uses adapter modules. Adapters can be small trainable layers that are inserted between pre-existing layers of a pre-trained model. During the fine-tuning process, the original parameters of the pre-trained model are typically frozen, and only the parameters of the adapters are updated.

[0208] In some implementations, example method 1900 can be implemented to execute parameter-efficient fine-tuning methods, such as Layerwise Optimization of Residuals (LoRA). LoRA can refine pre-trained models with minimal adjustments to the original parameters. Thiscan be achieved by introducing trainable low-rank matrices that modify the behavior of the pretrained weights without directly altering them. In some implementations, during fine-tuning, only these auxiliary matrices are updated, which significantly reduces the number of parameters that are trained.

[0209] An example fine-tuning approach includes reinforcement learning. Reinforcement learning can be based on user feedback on model performance during use.

[0210] Figure 20 is a block diagram of an example processing flow for using machine-learned model(s) 1 to process input(s) 2 to generate output(s) 3.

[0211] Machine-learned model(s) 1 can be or include one or multiple machine-learned models or model components. Example machine-learned models can include neural networks (e.g., deep neural networks). Example machine-learned models can include non-linear models or linear models. Example machine-learned models can use other architectures in lieu of or in addition to neural networks. Example machine-learned models can include decision tree based models, support vector machines, hidden Markov models, Bayesian networks, linear regression models, k-means clustering models, etc.

[0212] Machine-learned model(s) 1 can be or include, or otherwise be representative of any one or more of the machine-learned models described above with respect to the preceding figures. For example, machine-learned model(s) 1 can be or include, or otherwise be representative of any one or more of the models described herein, etc. Although various features, variations, and implementations described below are described with respect to machine-learned model(s) 1, it is to be understood that such features, variations, and implementations are to be understood as described with respect to each of the models described herein, etc., any other machine-learned component described herein.

[0213] Example neural networks can include feed-forward neural networks, recurrent neural networks (RNNs), including long short-term memory (LSTM) based recurrent neural networks, convolutional neural networks (CNNs), diffusion models, generative-adversarial networks, or other forms of neural networks. Example neural networks can be deep neural networks. Some example machine-learned models can leverage an attention mechanism such as self-attention. For example, some example machine-learned models can include multi-headed self-attention models.

[0214] Machine-learned model(s) 1 can include a single or multiple instances of the same model configured to operate on data from input(s) 2. Machine-learned model(s) 1 can include multiple different models or multiple different model portions configured to operate on data from input(s) 2.

[0215] Machine-learned model(s) 1 can include an ensemble of different models that can cooperatively interact to process data from input(s) 2. For example, a model ensemble can include multiple models that have different attributes (e.g., different architectures, trained with different recipes, etc.). The ensemble can output an overall output based on the individual outputs of the constituent models. In this manner, for instance, the diverse constituent models can work together to provide system-level robustness by effectively aggregating over individual strengths and weaknesses of any given model. The respective individual outputs can be combined in a weighted combination, using a voting or routing mechanism, or a learned output layer (e g., one or more feedforward or fully-connected layers).

[0216] Machine-learned model(s) 1 can employ a mixture-of-experts structure. See, e.g., Zhou et al., Mixture-of-Experts with Expert Choice Routing, ARXIV:2202.09368V2 (Oct. 14, 2022). For example, different portions of a model can learn (explicitly or implicitly) different expertise areas, with pathways through the model being selected by a learned routing mechanism that engages the appropriate expert for a given input (e.g., a given portion of an input, such as on a per-token basis). For example, a feedforward network can be sparsely activated for a given portion of an input based on an output of a routing mechanism that processes the portion of the input. In this manner, for instance, the group of activated weights can form an “expert” that is selected by the router. On each forward pass, only a subset of the total model weights may be engaged, thereby decreasing a quantity of operations performed for processing a given input compared to a densely activated model. In this manner, for instance, the expressive and interpretive power of a high-parameter-count model can be achieved with more computeefficient forward passes.

[0217] Input(s) 2 can generally include or otherwise represent various types of data. Input(s) 2 can include one type or many different types of data. Output(s) 3 can be data of the same type(s) or of different types of data as compared to input(s) 2. Output(s) 3 can include one type or many different types of data.

[0218] Example data types for input(s) 2 or output(s) 3 include natural language text data, software code data (e.g., source code, object code, machine code, or any other form of computer-readable instructions or programming languages), machine code data (e.g., binary code, assembly code, or other forms of machine-readable instructions that can be executed directly by a computer's central processing unit), assembly code data (e g., low-level programming languages that use symbolic representations of machine code instructions to program a processing unit), genetic data or other chemical or biochemical data, image data, audio data, audiovisual data, haptic data, biometric data, medical data, financial data, statistical data, geographical data, astronomical data, historical data, sensor data generally (e.g., digital or analog values, such as voltage or other absolute or relative level measurement values from a real or artificial input, such as from an audio sensor, light sensor, displacement sensor, etc.), and the like. Data can be raw or processed and can be in any format or schema.

[0219] In multimodal inputs 2 or outputs 3, example combinations of data types include image data and audio data, image data and natural language data, natural language data and software code data, image data and biometric data, sensor data and medical data, etc. It is to be understood that any combination of data types in an input 2 or an output 3 can be present.

[0220] An example input 2 can include one or multiple data types, such as the example data types noted above. An example output 3 can include one or multiple data types, such as the example data types noted above. The data type(s) of input 2 can be the same as or different from the data type(s) of output 3. It is to be understood that the example data types noted above are provided for illustrative purposes only. Data types contemplated within the scope of the present disclosure are not limited to those examples noted above.

[0221] Figure 21 is a block diagram of an example implementation of an example machine-learned model configured to process sequences of information. For instance, an example implementation of machine-learned model(s) 1 can include machine-learned sequence processing model(s) 4. An example system can pass input(s) 2 to sequence processing model(s) 4. Sequence processing model(s) 4 can include one or more machine-learned components. Sequence processing model(s) 4 can process the data from input(s) 2 to obtain an input sequence 5. Input sequence 5 can include one or more input elements 5-1, 5-2, . . . , 5-A7, etc. obtained from input(s) 2. Sequence processing model 4 can process input sequence 5 using prediction layer(s) 6 to generate an output sequence 7. Output sequence 7 can include one or more outputelements 7-1, 7-2, . . . , 7-N, etc. generated based on input sequence 5. The system can generate output(s) 3 based on output sequence 7.

[0222] Sequence processing model(s) 4 can include one or multiple machine-learned model components configured to ingest, generate, or otherwise reason over sequences of information. For example, some example sequence processing models are referred to as language models and can leverage language-based understandings across one or multiple modalities of input information. Sequence processing model(s) 4 can include relatively large models (e.g., more parameters, computationally expensive, etc.), which may be referred to as “Large Language Models” or LLMs. Sequence processing model(s) 4 can include relatively small models (e.g., fewer parameters, computationally lightweight, etc.), which may be referred to as “Small Language Models” or SLMs. Example language models include, for instance, models described in Gemma: Open Models Based on Gemini Research and Technology, GOOGLE, https: / / arxiv.org / abs / 2403.08295; Gemma 2: Improving Open Language Models at a Practical Size, GOOGLE, https: / / arxiv.org / abs / 2408.00118.

[0223] Sequence processing model(s) 4 can process one or multiple types of data simultaneously. Variations of language models that can perform joint vision and language tasks may be referred to as “Vision-Language Models,” or VLMs. Example VLMs include models described in PaliGemma: A versatile 3B VLM for transfer, GOOGLE, https: / / arxiv.org / abs / 2407.07726; PaliGemma 2: A Family of Versatile VLMs for Transfer, GOOGLE, https: / / arxiv.org / abs / 2412.03555; Flamingo: a Visual Language Model for Few-Shot Learning, GOOGLE, https: / / arxiv.org / abs / 2204.14198; PaLI: A Jointly-Scaled Multilingual Language-Image Model, GOOGLE, https: / / arxiv.org / abs / 2209.06794.

[0224] Sequence processing model(s) 4 can be multimodal. Example multimodal sequence processing models include, for instance, models described in Gemini: A Family of Highly Capable Multimodal Models, GOOGLE, https: / / arxiv.org / abs / 2312.11805; Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context, GOOGLE, https: / / arxiv.org / abs / 2403.05530.

[0225] Other example sequence processing models can operate to generate outputs or receive inputs in specific domains, such as image domains, see, e.g., Dosovitskiy et al., An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale, ARXIV:2010.11929v2 (Jun. 3, 2021), audio domains, see, e.g., Agostinelli et al., MusicL M: Generating Music FromText, ARXTV:2301.11325vl (Jan. 26, 2023), biochemical domains, see, e.g. , Jumper et al., Highly accurate protein structure prediction with AlphaFold, 596 Nature 583 (Aug. 26, 2021), by way of example.

[0226] In general, sequence processing model(s) 4 can obtain input sequence 5 using data from input(s) 2. For instance, input sequence 5 can include a representation of data from input(s) 2 in a format understood by sequence processing model(s) 4. One or more machine-learned components of sequence processing model(s) 4 can ingest the data from input(s) 2, parse the data into pieces compatible with the processing architectures of sequence processing model(s) 4 (e.g., via “tokenization”), and project the pieces into an input space associated with prediction layer(s) 6 (e.g., via “embedding”).

[0227] Sequence processing model(s) 4 can ingest the data from input(s) 2 and parse the data into a sequence of elements to obtain input sequence 5. For example, a portion of input data from input(s) 2 can be broken down into pieces that collectively represent the content of the portion of the input data. The pieces can provide the elements of the sequence.

[0228] Elements 5-1, 5-2, . . . , 5-AL can represent, in some cases, building blocks for capturing or expressing meaningful information in a particular data domain. For instance, the elements can describe “atomic units” across one or more domains. For example, for textual input source(s), the elements can correspond to groups of one or more words or sub-word components, such as sets of one or more characters.

[0229] For example, elements 5-1, 5-2, . . . , 5-M can represent tokens obtained using a tokenizer. For instance, a tokenizer can process a given portion of an input source and output a series of tokens (e.g., corresponding to input elements 5-1, 5-2, . . . , 5-M) that represent the portion of the input source. Various approaches to tokenization can be used. For instance, textual input source(s) can be tokenized using a byte-pair encoding (BPE) technique. See, e.g., Kudo et al., SentencePiece: A simple and language independent subword tokenizer and detokenizer for Neural Text Processing, PROCEEDINGS OF THE 2018 CONFERENCE ON EMPIRICAL METHODS IN NATURAL L NGUAGE PROCESSING (System Demonstrations), pages 66-71 (October 31-November 4, 2018), https: / / aclanthology.org / D18-2012.pdf. Image-based input source(s) can be tokenized by extracting and serializing patches from an image.

[0230] In general, arbitrary data types can be serialized and processed into input sequence 5. It is to be understood that element(s) 5-1, 5-2, . . . , 5-M depicted in Figure 21 can be the tokens or can be the embedded representations thereof.

[0231] Prediction layer(s) 6 can predict one or more output elements 7-1, 7-2, . . . , 7-N based on the input elements. Prediction layer(s) 6 can include one or more machine-learned model architectures, such as one or more layers of learned parameters that manipulate and transform the input(s) to extract higher-order meaning from, and relationships between, input element(s) 5-1, 5-2, . . . , 5-M. In this manner, for instance, example prediction layer(s) 6 can predict new output element(s) in view of the context provided by input sequence 5.

[0232] Prediction layer(s) 6 can evaluate associations between portions of input sequence 5 and a particular output element. These associations can inform a prediction of the likelihood that a particular output follows the input context. For example, consider the textual snippet, “The carpenter’s toolbox was small and heavy. It was full of .” Example prediction layer(s) 6 can identify that “It” refers back to “toolbox” by determining a relationship between the respective embeddings. Example prediction layer(s) 6 can also link “It” to the attributes of the toolbox, such as “small” and “heavy.” Based on these associations, prediction layer(s) 6 can, for instance, assign a higher probability to the word “nails” than to the word “sawdust.”

[0233] A transformer is an example architecture that can be used in prediction layer(s) 4. See, e.g., Vaswani et al., Attention Is All You Need, ARXIV: 1706.03762v7 (Aug. 2, 2023). A transformer is an example of a machine-learned model architecture that uses an attention mechanism to compute associations between items within a context window. The context window can include a sequence that contains input sequence 5 and potentially one or more output element(s) 7-1, 7-2, . . . , 7-N. A transformer block can include one or more attention layer(s) and one or more post-attention layer(s) (e.g., feedforward layer(s), such as a multi-layer perceptron).

[0234] Prediction layer(s) 6 can include other machine-learned model architectures in addition to or in lieu of transformer-based architectures. For example, recurrent neural networks (RNNs) and long short-term memory (LSTM) models can also be used, as well as convolutional neural networks (CNNs). In general, prediction layer(s) 6 can leverage various kinds of artificial neural networks that can understand or generate sequences of information.

[0235] Output sequence 7 can include or otherwise represent the same or different data types as input sequence 5. For instance, input sequence 5 can represent textual data, and output sequence 7 can represent textual data. Input sequence 5 can represent image, audio, or audiovisual data, and output sequence 7 can represent textual data (e.g., describing the image, audio, or audiovisual data). It is to be understood that prediction layer(s) 6, and any other interstitial model components of sequence processing model (s) 4, can be configured to receive a variety of data types in input sequence(s) 5 and output a variety of data types in output sequence(s) 7.

[0236] Output sequence 7 can have various relationships to input sequence 5. Output sequence 7 can be a continuation of input sequence 5. Output sequence 7 can be complementary to input sequence 5. Output sequence 7 can translate, transform, augment, or otherwise modify input sequence 5. Output sequence 7 can answer, evaluate, confirm, or otherwise respond to input sequence 5. Output sequence 7 can implement (or describe instructions for implementing) an instruction provided via input sequence 5.

[0237] Output sequence 7 can be generated autoregressively. For instance, for some applications, an output of one or more prediction layer(s) 6 can be passed through one or more output layers (e.g., softmax layer) to obtain a probability distribution over an output vocabulary (e.g., a textual or symbolic vocabulary) conditioned on a set of input elements in a context window. In this manner, for instance, output sequence 7 can be autoregressively generated by sampling a likely next output element, adding that element to the context window, and regenerating the probability distribution based on the updated context window, and sampling a likely next output element, and so forth.

[0238] Output sequence 7 can also be generated non-autoregressively. For instance, multiple output elements of output sequence 7 can be predicted together without explicit sequential conditioning on each other. See, e.g., Saharia et al., Non-Autoregressive Machine Translation with Latent Alignments, ARXIV:2004.07437V3 (NOV. 16, 2020).

[0239] Output sequence 7 can include one or multiple portions or elements. In an example content generation configuration, output sequence 7 can include multiple elements corresponding to multiple portions of a generated output sequence (e.g., a textual sentence, values of a discretized waveform, computer code, etc.). In an example classification configuration, output sequence 7 can include a single element associated with a classificationoutput. For instance, an output “vocabulary” can include a set of classes into which an input sequence is to be classified. For instance, a vision transformer block can pass latent state information to a multilayer perceptron that outputs a likely class value associated with an input image.

[0240] Figure 22 is a block diagram of an example technique for populating an example input sequence 8. Input sequence 8 can include various functional elements that form part of the model infrastructure, such as an element 8-0 obtained from a task indicator 9 that signals to any model(s) that process input sequence 8 that a particular task is being performed (e.g., to help adapt a performance of the model(s) to that particular task). Input sequence 8 can include various data elements from different data modalities. For instance, an input modality 10-1 can include one modality of data. A data-to-sequence model 11-1 can process data from input modality 10-1 to project the data into a format compatible with input sequence 8 (e.g., one or more vectors dimensioned according to the dimensions of input sequence 8) to obtain elements 8-1, 8-2, 8-3. Another input modality 10-2 can include a different modality of data. A data-to-sequence model 11-2 can project data from input modality 10-2 into a format compatible with input sequence 8 to obtain elements 8-4, 8-5, 8-6. Another input modality 10-3 can include yet another different modality of data. A data-to-sequence model 11-3 can project data from input modality 10-3 into a format compatible with input sequence 8 to obtain elements 8-7, 8-8, 8-9.

[0241] Input sequence 8 can be the same as or different from input sequence 5. Input sequence 8 can be a multimodal input sequence that contains elements that represent data from different modalities using a common dimensional representation. For instance, an embedding space can have P dimensions. Input sequence 8 can be configured to contain a plurality of elements that have P dimensions. In this manner, for instance, example implementations can facilitate information extraction and reasoning across diverse data modalities by projecting data into elements in the same embedding space for comparison, combination, or other computations therebetween.

[0242] For example, elements 8-0, . . . , 8-9 can indicate particular locations within a multidimensional embedding space. Some elements can map to a set of discrete locations in the embedding space. For instance, elements that correspond to discrete members of a predetermined vocabulary of tokens can map to discrete locations in the embedding space that are associated with those tokens. Other elements can be continuously distributed across the embedding space.For instance, some data types can be broken down into continuously defined portions (e.g., image patches) that can be described using continuously distributed locations within the embedding space.

[0243] In some implementations, the expressive power of the embedding space may not be limited to meanings associated with any particular set of tokens or other building blocks. For example, a continuous embedding space can encode a spectrum of high-order information. An individual piece of information (e.g., a token) can map to a particular point in that space: for instance, a token for the word “dog” can be projected to an embedded value that points to a particular location in the embedding space associated with canine-related information. Similarly, an image patch of an image of a dog on grass can also be projected into the embedding space. In some implementations, the projection of the image of the dog can be similar to the projection of the word “dog” while also having similarity to a projection of the word “grass,” while potentially being different from both. In some implementations, the projection of the image patch may not exactly align with any single projection of a single word. In some implementations, the projection of the image patch can align with a combination of the projections of the words “dog” and “grass.” In this manner, for instance, a high-order embedding space can encode information that can be independent of data modalities in which the information is expressed.

[0244] Task indicator 9 can include a model or model component configured to identify a task being performed and inject, into input sequence 8, an input value represented by element 8-0 that signals which task is being performed. For instance, the input value can be provided as a data type associated with an input modality and projected along with that input modality (e.g., the input value can be a textual task label that is embedded along with other textual data in the input; the input value can be a pixel -based representation of a task that is embedded along with other image data in the input; etc ). The input value can be provided as a data type that differs from or is at least independent from other input(s). For instance, the input value represented by element 8-0 can be learned within a continuous embedding space.

[0245] Input modalities 10-1, 10-2, and 10-3 can be associated with various different data types (e.g., as described above with respect to input(s) 2 and output(s) 3).

[0246] Data-to-sequence models 11-1, 11-2, and 11-3 can be the same or different from each other. Data-to-sequence models 11-1, 11-2, and 11-3 can be adapted to each respective input modality 10-1, 10-2, and 10-3. For example, a textual data-to-sequence model cansubdivide a portion of input text and project the subdivisions into element(s) in input sequence 8 (e.g., elements 8-1, 8-2, 8-3, etc.). An image data-to-sequence model can subdivide an input image and project the subdivisions into element(s) in input sequence 8 (e.g., elements 8-4, 8-5, 8-6, etc.). An arbitrary data type data-to-sequence model can subdivide an input of that arbitrary data type and project the subdivisions into element(s) in input sequence 8 (e.g., elements 8-7, 8-8, 8-9, etc.).

[0247] Data-to-sequence models 11-1, 11-2, and 11-3 can form part of machine-learned sequence processing model(s) 4. Data-to-sequence models 11-1, 11-2, and 11-3 can be jointly trained with or trained independently from machine-learned sequence processing model(s) 4. Data-to-sequence models 11-1, 11-2, and 11-3 can be trained end-to-end with machine-learned sequence processing model(s) 4.

[0248] Figure 23 is a block diagram of an example model development platform 12 that can facilitate creation, adaptation, and refinement of example machine-learned models (e.g., machine-learned model(s) 1, sequence processing model(s) 4, etc.). Model development platform 12 can provide a number of different toolkits that developer systems can employ in the development of new or adapted machine-learned models.

[0249] Model development platform 12 can provide one or more model libraries 13 containing building blocks for new models. Model libraries 13 can include one or more pretrained foundational models 13-1, which can provide a backbone of processing power across various tasks. Model libraries 13 can include one or more pre-trained expert models 13-2, which can be focused on performance in particular domains of expertise. Model libraries 13 can include various model primitives 13-3, which can provide low-level architectures or components (optionally pre-trained), which can be assembled in various arrangements as desired. Model primitives 13-3 can include a library of pre-trained adapters or LoRA modules that can adapt a baseline foundational model to align its outputs with a desired performance profde, augment model capabilities (e.g., to adapt to a different input modality, etc.), and the like.

[0250] Model development platform 12 can receive selections of various model components 14. Model development platform 12 can pass selected model components 14 to a workbench 15 that combines selected model components 14 into a development model 16.

[0251] Workbench 15 can facilitate further refinement and adaptation of development model 16 by leveraging a number of different toolkits integrated with model developmentplatform 12. For example, workbench 1 can facilitate alignment of the development model 16 with a desired performance profde on various tasks using a model alignment toolkit 17.

[0252] Model alignment toolkit 17 can provide a number of tools for causing development model 16 to generate outputs aligned with desired behavioral characteristics.Alignment can include increasing the accuracy, precision, recall, etc. of model outputs.Alignment can include enforcing output styles, schema, or other preferential characteristics of model outputs. Alignment can be general or domain-specific. For instance, a pre-trained foundational model 13-1 can begin with an initial level of performance across multiple domains. Alignment of the pre-trained foundational model 13-1 can include improving a performance in a particular domain of information or tasks (e.g., even at the expense of performance in another domain of information or tasks).

[0253] Model alignment toolkit 17 can integrate one or more dataset(s) 17-1 for aligning development model 16. Curated dataset(s) 17-1 can include labeled or unlabeled training data. Dataset(s) 17-1 can be obtained from public domain datasets. Dataset(s) 17-1 can be obtained from private datasets associated with one or more developer system(s) for the alignment of bespoke machine-learned model(s) customized for private use-cases.

[0254] Pre-training pipelines 17-2 can include a machine-learned model training workflow configured to update development model 16 over large-scale, potentially noisy datasets. For example, pre-training can leverage unsupervised learning techniques (e.g., denoising, etc.) to process large numbers of training instances to update model parameters from an initialized state and achieve a desired baseline performance. Pre-training pipelines 17-2 can leverage unlabeled datasets in dataset(s) 17-1 to perform pre-training. Workbench 15 can implement a pre-training pipeline 17-2 to pre-train development model 16.

[0255] Fine-tuning pipelines 17-3 can include a machine-learned model training workflow configured to refine the model parameters of development model 16 with higher-quality data. Fine-tuning pipelines 17-3 can update development model 16 by conducting supervised training with labeled dataset(s) in dataset(s) 17-1. Fine-tuning pipelines 17-3 can update development model 16 by conducting reinforcement learning using reward signals from user feedback signals. Workbench 15 can implement a fine-tuning pipeline 17-3 to fine-tune development model 16.

[0256] Prompt libraries 17-4 can include sets of inputs configured to induce behavior aligned with desired performance criteria. Prompt libraries 17-4 can include few-shot prompts (e.g., inputs providing examples of desired model outputs for prepending to a desired runtime query), chain-of-thought prompts (e.g., inputs providing step-by-step reasoning within the exemplars to facilitate thorough reasoning by the model), and the like.

[0257] Example prompts can be retrieved from an available repository of prompt libraries 17-4. Example prompts can be contributed by one or more developer systems using workbench 15.

[0258] In some implementations, pre-trained or fine-tuned models can achieve satisfactory performance without exemplars in the inputs. For instance, zero-shot prompts can include inputs that lack exemplars. Zero-shot prompts can be within a domain within a training dataset or outside of the training domain(s).

[0259] Prompt libraries 17-4 can include one or more prompt engineering tools. Prompt engineering tools can provide workflows for retrieving or learning optimized prompt values. Prompt engineering tools can facilitate directly learning prompt values (e.g., input element values) based on one or more training iterations. Workbench 15 can implement prompt engineering tools in development model 16.

[0260] Prompt libraries 17-4 can include pipelines for prompt generation. For example, inputs can be generated using development model 16 itself or other machine-learned models. In this manner, for instance, a first model can process information about a task and output an input for a second model to process in order to perform a step of the task. The second model can be the same as or different from the first model. Workbench 15 can implement prompt generation pipelines in development model 16.

[0261] Prompt libraries 17-4 can include pipelines for context injection. For instance, a performance of development model 16 on a particular task can improve if provided with additional context for performing the task. Prompt libraries 17-4 can include software components configured to identify desired context, retrieve the context from an external source (e.g., a database, a sensor, etc.), and add the context to the input prompt. Workbench 15 can implement context injection pipelines in development model 16.

[0262] Although various training examples described herein with respect to model development platform 12 refer to “pre-training” and “fine-tuning,” it is to be understood thatmodel alignment toolkit 17 can generally support a wide variety of training techniques adapted for training a wide variety of machine-learned models. Example training techniques can correspond to the example training method 1900 described above.

[0263] Model development platform 12 can include a model plugin toolkit 18. Model plugin toolkit 18 can include a variety of tools configured for augmenting the functionality of a machine-learned model by integrating the machine-learned model with other systems, devices, and software components. For instance, a machine-learned model can use tools to increase performance quality where appropriate. For instance, deterministic tasks can be offloaded to dedicated tools in lieu of probabilistically performing the task with an increased risk of error. For instance, instead of autoregressively predicting the solution to a system of equations, a machine-learned model can recognize a tool to call for obtaining the solution and pass the system of equations to the appropriate tool. The tool can be a traditional system of equations solver that can operate deterministically to resolve the system of equations. The output of the tool can be returned in response to the original query. In this manner, tool use can allow some example models to focus on the strengths of machine-learned models — e.g., understanding an intent in an unstructured request for a task — while augmenting the performance of the model by offloading certain tasks to a more focused tool for rote application of deterministic algorithms to a well-defined problem.

[0264] Model plugin toolkit 18 can include validation tools 18-1. Validation tools 18-1 can include tools that can parse and confirm output(s) of a machine-learned model. Validation tools 18-1 can include engineered heuristics that establish certain thresholds applied to model outputs. For example, validation tools 18-1 can ground the outputs of machine-learned models to structured data sources (e.g., to mitigate “hallucinations”).

[0265] Model plugin toolkit 18 can include tooling packages 18-2 for implementing one or more tools that can include scripts or other executable code that can be executed alongside development model 16. Tooling packages 18-2 can include one or more inputs configured to cause machine-learned model(s) to implement the tools (e.g., few-shot prompts that induce a model to output tool calls in the proper syntax, etc.). Tooling packages 18-2 can include, for instance, fine-tuning training data for training a model to use a tool.

[0266] Model plugin toolkit 18 can include interfaces for calling external application programming interfaces (APIs) 18-3. For instance, in addition to or in lieu of implementing toolcalls or tool code directly with development model 16, development model 16 can be aligned to output instructions that initiate API calls to send or obtain data via external systems.

[0267] Model plugin toolkit 18 can integrate with prompt libraries 17-4 to build a catalog of available tools for use with development model 16. For instance, a model can receive, in an input, a catalog of available tools, and the model can generate an output that selects a tool from the available tools and initiates a tool call for using the tool.

[0268] Model development platform 12 can include a computational optimization toolkit 19 for optimizing a computational performance of development model 16. For instance, tools for model compression 19-1 can allow development model 16 to be reduced in size while maintaining a desired level of performance. For instance, model compression 19-1 can include quantization workflows, weight pruning and sparsification techniques, etc. Tools for hardware acceleration 19-2 can facilitate the configuration of the model storage and execution formats to operate optimally on different hardware resources. For instance, hardware acceleration 19-2 can include tools for optimally sharding models for distributed processing over multiple processing units for increased bandwidth, lower unified memory requirements, etc. Tools for distillation 19-3 can provide for the training of lighter-weight models based on the knowledge encoded in development model 16. For instance, development model 16 can be a highly performant, large machine-learned model optimized using model development platform 12. To obtain a lightweight model for running in resource-constrained environments, a smaller model can be a “student model” that learns to imitate development model 16 as a “teacher model.” In this manner, for instance, the investment in learning the parameters and configurations of development model 16 can be efficiently transferred to a smaller model for more efficient inference.

[0269] Workbench 15 can implement one, multiple, or none of the toolkits implemented in model development platform 12. Workbench 15 can output an output model 20 based on development model 16. Output model 20 can be a deployment version of development model 16. Output model 20 can be a development or training checkpoint of development model 16. Output model 20 can be a distilled, compressed, or otherwise optimized version of development model 16.

[0270] Figure 24 is a block diagram of an example training flow for training a machine-learned development model 16. One or more portion(s) of the example training flow can beimplemented by a computing system that includes one or more computing devices such as, for example, computing systems described with reference to the other figures. Each respective portion of the example training flow can be performed by any (or any combination) of one or more computing devices. Moreover, one or more portion(s) of the example training flow can be implemented on the hardware components of the device(s) described herein, for example, to train one or more systems or models. Figure 24 depicts elements performed in a particular order for purposes of illustration and discussion. Those of ordinary skill in the art, using the disclosures provided herein, will understand that the elements of any of the methods discussed herein can be adapted, rearranged, expanded, omitted, combined, or modified in various ways without deviating from the scope of the present disclosure. Figure 24 is described with reference to elements / terms described with respect to other systems and figures for exemplary illustrated purposes and is not meant to be limiting. One or more portions of the example training flow can be performed additionally, or alternatively, by other systems.

[0271] Initially, development model 16 can persist in an initial state as an initialized model 21. Development model 16 can be initialized with weight values. Initial weight values can be random or based on an initialization schema. Initial weight values can be based on prior pretraining for the same or for a different model.

[0272] Initialized model 21 can undergo pre-training in a pre-training stage 22. Pretraining stage 22 can be implemented using one or more pre-training pipelines 17-2 over data from dataset(s) 17-1. Pre-training can be omitted, for example, if initialized model 21 is already pre-trained (e.g., development model 16 contains, is, or is based on a pre-trained foundational model or an expert model).

[0273] Pre-trained model 23 can then be a new version of development model 16, which can persist as development model 16 or as a new development model. Pre-trained model 23 can be the initial state if development model 16 was already pre-trained. Pre-trained model 23 can undergo fine-tuning in a fine-tuning stage 24. Fine-tuning stage 24 can be implemented using one or more fine-tuning pipelines 17-3 over data from dataset(s) 17-1. Fine-tuning can be omitted, for example, if a pre-trained model has satisfactory performance, if the model was already fine-tuned, or if other tuning approaches are preferred.

[0274] Fine-tuned model 29 can then be a new version of development model 16, which can persist as development model 16 or as a new development model. Fine-tuned model 29 canbe the initial state if development model 16 was already fine-tuned. Fine-tuned model 29 can undergo refinement with user feedback 26. For instance, refinement with user feedback 26 can include reinforcement learning, optionally based on human feedback from human users of finetuned model 25. As reinforcement learning can be a form of fine-tuning, it is to be understood that fine-tuning stage 24 can subsume the stage for refining with user feedback 26. Refinement with user feedback 26 can produce a refined model 27. Refined model 27 can be output to downstream system(s) 28 for deployment or further development.

[0275] In some implementations, computational optimization operations can be applied before, during, or after each stage. For instance, initialized model 21 can undergo computational optimization 29-1 (e.g., using computational optimization toolkit 19) before pre-training stage 22. Pre-trained model 23 can undergo computational optimization 29-2 (e.g., using computational optimization toolkit 19) before fine-tuning stage 24. Fine-tuned model 25 can undergo computational optimization 29-3 (e.g., using computational optimization toolkit 19) before refinement with user feedback 26. Refined model 27 can undergo computational optimization 29-4 (e.g., using computational optimization toolkit 19) before output to downstream system(s) 28. Computational optimization(s) 29-1, . . . , 29-4 can all be the same, all be different, or include at least some different optimization techniques.

[0276] Figure 25 is a block diagram of an inference system for operating one or more machine-learned model(s) 1 to perform inference (e.g., for training, for deployment, etc ). A model host 31 can receive machine-learned model(s) 1. Model host 31 can host one or more model instance(s) 31-1, which can be one or multiple instances of one or multiple models. Model host 31 can host model instance(s) 31-1 using available compute resources 31-2 associated with model host 31.

[0277] Model host 31 can perform inference on behalf of one or more client(s) 32.Client(s) 32 can transmit an input request 33 to model host 31. Using input request 33, model host 31 can obtain input(s) 2 for input to machine-learned model(s) 1. Machine-learned model(s) 1 can process input(s) 2 to generate output(s) 3. Using output(s) 3, model host 31 can return an output payload 34 for responding to input request 33 from client(s) 32. Output payload 34 can include or be based on output(s) 3.

[0278] Model host 31 can leverage various other resources and tools to augment the inference task. For instance, model host 31 can communicate with tool interfaces 35 to facilitatetool use by model instance(s) 31-1. Tool interfaces 35 can include local or remote APIs. Tool interfaces 35 can include integrated scripts or other software functionality. Model host 31 can engage online learning interface(s) 36 to facilitate ongoing improvements to machine-learned model(s) 1. For instance, online learning interface(s) 36 can be used within reinforcement learning loops to retrieve user feedback on inferences served by model host 31. Model host 31 can access runtime data source(s) 37 for augmenting input(s) 2 with additional contextual information. For instance, runtime data source(s) 37 can include a knowledge graph 37-1 that facilitates structured information retrieval for information associated with input request(s) 33 (e.g., a search engine service). Runtime data source(s) 37 can include public or private, external or local database(s) 37-2 that can store information associated with input request(s) 33 for augmenting input(s) 2. Runtime data source(s) 37 can include account data 37-3 which can be retrieved in association with a user account corresponding to a client 32 for customizing the behavior of model host 31 accordingly.

[0279] Model host 31 can be implemented by one or multiple computing devices or systems. Client(s) 2 can be implemented by one or multiple computing devices or systems, which can include computing devices or systems shared with model host 31.

[0280] For example, model host 31 can operate on a server system that provides a machine-learning service to client device(s) that operate client(s) 32 (e.g., over a local or wide-area network). Client device(s) can be end-user devices used by individuals. Client device(s) can be server systems that operate client(s) 32 to provide various functionality as a service to downstream end-user devices.

[0281] In some implementations, model host 31 can operate on the same device or system as client(s) 32. Model host 31 can be a machine-learning service that runs on-device to provide machine-learning functionality to one or multiple applications operating on a client device, which can include an application implementing client(s) 32. Model host 31 can be a part of the same application as client(s) 32. For instance, model host 31 can be a subroutine or method implemented by one part of an application, and client(s) 32 can be another subroutine or method that engages model host 31 to perform inference functions within the application. It is to be understood that model host 31 and client(s) 32 can have various different configurations.

[0282] Model instance(s) 31-1 can include one or more machine-learned models that are available for performing inference. Model instance(s) 31-1 can include weights or other modelcomponents that are stored in persistent storage, temporarily cached, or loaded into high-speed memory. Model instance(s) 31-1 can include multiple instance(s) of the same model (e.g., for parallel execution of more requests on the same model). Model instance(s) 31-1 can include instance(s) of different model(s). Model instance(s) 31-1 can include cached intermediate states of active or inactive model(s) used to accelerate inference of those models. For instance, an inference session with a particular model may generate significant amounts of computational results that can be re-used for future inference runs (e.g., using a KV cache for transformer-based models). These computational results can be saved in association with that inference session so that session can be executed more efficiently when resumed.

[0283] Compute resource(s) 31-2 can include one or more processors (central processing units, graphical processing units, tensor processing units, machine-learning accelerators, etc.) connected to one or more memory devices. Compute resource(s) 31-2 can include a dynamic pool of available resources shared with other processes. Compute resource(s) 31-2 can include memory devices large enough to fit an entire model instance in a single memory instance.Compute resource(s) 31-2 can also shard model instance(s) across multiple memory devices (e.g., using data parallelization or tensor parallelization, etc.). This can be done to increase parallelization or to execute a large model using multiple memory devices which individually might not be able to fit the entire model into memory.

[0284] Input request 33 can include data for input(s) 2. Model host 31 can process input request 33 to obtain input(s) 2. Input(s) 2 can be obtained directly from input request 33 or can be retrieved using input request 33. Input request 33 can be submitted to model host 31 via an API.

[0285] Model host 31 can perform inference over batches of input requests 33 in parallel. For instance, a model instance 31-1 can be configured with an input structure that has a batch dimension. Separate input(s) 2 can be distributed across the batch dimension (e.g., rows of an array). The separate input(s) 2 can include completely different contexts. The separate input(s) 2 can be multiple inference steps of the same task. The separate input(s) 2 can be staggered in an input structure, such that any given inference cycle can be operating on different portions of the respective input(s) 2. In this manner, for instance, model host 31 can perform inference on the batch in parallel, such that output(s) 3 can also contain the batch dimension and return theinference results for the batched input(s) 2 in parallel. In this manner, for instance, batches of input request(s) 33 can be processed in parallel for higher throughput of output payload(s) 34.

[0286] Output payload 34 can include or be based on output(s) 3 from machine-learned model(s) 1. Model host 31 can process output(s) 3 to obtain output payload 34. This can include chaining multiple rounds of inference (e.g., iteratively, recursively, across the same model(s) or different model(s)) to arrive at a final output for a task to be returned in output payload 34.Output payload 34 can be transmitted to client(s) 32 via an API.

[0287] Online learning interface(s) 36 can facilitate reinforcement learning of machine-learned model(s) 1. Online learning interface(s) 36 can facilitate reinforcement learning with human feedback (RLHF). Online learning interface(s) 36 can facilitate federated learning of machine-learned model(s) 1.

[0288] Model host 31 can access a library of pre-trained adapters or LoRA modules that can adapt a baseline model to align its outputs with a desired performance profile, augment model capabilities (e.g., to adapt to a different input modality, etc.), and the like. For instance, model host 31 can receive an input request to load a customized model, and model host 31 can retrieve one or more components to adapt a baseline model to the custom profile. Model host 31 can determine that a particular functionality is needed for a particular task (e.g., based on an output of a model that preprocesses an input) and retrieve a pre-trained component accordingly.

[0289] Model host 31 can execute machine-learned model(s) 1 to perform inference for various tasks using various types of data. For example, various different input(s) 2 and output(s) 3 can be used for various different tasks. In some implementations, input(s) 2 can be or otherwise represent image data. Machine-learned model(s) 1 can process the image data to generate an output. As an example, machine-learned model(s) 1 can process the image data to generate an image recognition output (e.g., a recognition of the image data, a latent embedding of the image data, an encoded representation of the image data, a hash of the image data, etc.). As another example, machine-learned model(s) 1 can process the image data to generate an image segmentation output. As another example, machine-learned model(s) 1 can process the image data to generate an image classification output. As another example, machine-learned model(s) 1 can process the image data to generate an image data modification output (e.g., an alteration of the image data, etc.). As another example, machine-learned model(s) 1 can process the image data to generate an encoded image data output (e.g., an encoded and / or compressedrepresentation of the image data, etc.). As another example, machine-learned model(s) 1 can process the image data to generate an upscaled image data output. As another example, machine-learned model(s) 1 can process the image data to generate a prediction output.

[0290] In some implementations, the task is a computer vision task. In some cases, input(s) 2 includes pixel data for one or more images and the task is an image processing task. For example, the image processing task can be image classification, where the output is a set of scores, each score corresponding to a different object class and representing the likelihood that the one or more images depict an object belonging to the object class. The image processing task may be object detection, where the image processing output identifies one or more regions in the one or more images and, for each region, a likelihood that region depicts an object of interest. As another example, the image processing task can be image segmentation, where the image processing output defines, for each pixel in the one or more images, a respective likelihood for each category in a predetermined set of categories. For example, the set of categories can be foreground and background. As another example, the set of categories can be object classes. As another example, the image processing task can be depth estimation, where the image processing output defines, for each pixel in the one or more images, a respective depth value. As another example, the image processing task can be motion estimation, where the network input includes multiple images, and the image processing output defines, for each pixel of one of the input images, a motion of the scene depicted at the pixel between the images in the network input.

[0291] In some implementations, input(s) 2 can be or otherwise represent natural language data. Machine-learned model(s) 1 can process the natural language data to generate an output. As an example, machine-learned model(s) 1 can process the natural language data to generate a language encoding output. As another example, machine-learned model(s) 1 can process the natural language data to generate a latent text embedding output. As another example, machine-learned model(s) 1 can process the natural language data to generate a translation output. As another example, machine-learned model(s) 1 can process the natural language data to generate a classification output. As another example, machine-learned model(s) 1 can process the natural language data to generate a textual segmentation output. As another example, machine-learned model(s) 1 can process the natural language data to generate a semantic intent output. As another example, machine-learned model(s) 1 can process the natural language data to generate an upscaled text or natural language output (e.g., text or naturallanguage data that is higher quality than the input text or natural language, etc.). As another example, machine-learned model(s) 1 can process the natural language data to generate a prediction output (e.g., one or more predicted next portions of natural language content).

[0292] In some implementations, input(s) 2 can be or otherwise represent speech data (e.g., data describing spoken natural language, such as audio data, textual data, etc.). Machine-learned model(s) 1 can process the speech data to generate an output. As an example, machine-learned model(s) 1 can process the speech data to generate a speech recognition output. As another example, machine-learned model(s) 1 can process the speech data to generate a speech translation output. As another example, machine-learned model(s) 1 can process the speech data to generate a latent embedding output. As another example, machine-learned model(s) 1 can process the speech data to generate an encoded speech output (e.g., an encoded and / or compressed representation of the speech data, etc ). As another example, machine-learned model(s) 1 can process the speech data to generate an upscaled speech output (e.g., speech data that is higher quality than the input speech data, etc.). As another example, machine-learned model(s) 1 can process the speech data to generate a textual representation output (e.g., a textual representation of the input speech data, etc.). As another example, machine-learned model(s) 1 can process the speech data to generate a prediction output.

[0293] In some implementations, input(s) 2 can be or otherwise represent latent encoding data (e.g., a latent space representation of an input, etc.). Machine-learned model(s) 1 can process the latent encoding data to generate an output. As an example, machine-learned model(s) 1 can process the latent encoding data to generate a recognition output. As another example, machine-learned model(s) 1 can process the latent encoding data to generate a reconstruction output. As another example, machine-learned model(s) 1 can process the latent encoding data to generate a search output. As another example, machine-learned model(s) 1 can process the latent encoding data to generate a reclustering output. As another example, machine-learned model(s) 1 can process the latent encoding data to generate a prediction output.

[0294] In some implementations, input(s) 2 can be or otherwise represent statistical data. Statistical data can be, represent, or otherwise include data computed and / or calculated from some other data source. Machine-learned model(s) 1 can process the statistical data to generate an output. As an example, machine-learned model(s) 1 can process the statistical data to generate a recognition output. As another example, machine-learned model(s) 1 can process the statisticaldata to generate a prediction output. As another example, machine-learned model(s) 1 can process the statistical data to generate a classification output. As another example, machine-learned model(s) 1 can process the statistical data to generate a segmentation output. As another example, machine-learned model(s) 1 can process the statistical data to generate a visualization output. As another example, machine-learned model(s) 1 can process the statistical data to generate a diagnostic output.

[0295] In some implementations, input(s) 2 can be or otherwise represent sensor data. Machine-learned model(s) 1 can process the sensor data to generate an output. As an example, machine-learned model(s) 1 can process the sensor data to generate a recognition output. As another example, machine-learned model(s) 1 can process the sensor data to generate a prediction output. As another example, machine-learned model(s) 1 can process the sensor data to generate a classification output. As another example, machine-learned model(s) 1 can process the sensor data to generate a segmentation output. As another example, machine-learned model(s) 1 can process the sensor data to generate a visualization output. As another example, machine-learned model(s) 1 can process the sensor data to generate a diagnostic output. As another example, machine-learned model(s) 1 can process the sensor data to generate a detection output.

[0296] In some implementations, machine-learned model(s) 1 can be configured to perform a task that includes encoding input data for reliable and / or efficient transmission or storage (and / or corresponding decoding). For example, the task may be an audio compression task. The input may include audio data and the output may comprise compressed audio data. In another example, the input includes visual data (e.g. one or more images or videos), the output comprises compressed visual data, and the task is a visual data compression task. In another example, the task may comprise generating an embedding for input data (e.g. input audio or visual data). In some cases, the input includes audio data representing a spoken utterance and the task is a speech recognition task. The output may comprise a text output which is mapped to the spoken utterance. In some cases, the task comprises encrypting or decrypting input data. In some cases, the task comprises a microprocessor performance task, such as branch prediction or memory address translation.

[0297] In some implementations, the task is a generative task, and machine-learned model(s) 1 can be configured to output content generated in view of input(s) 2. For instance,input(s) 2 can be or otherwise represent data of one or more modalities that encodes context for generating additional content.

[0298] In some implementations, the task can be a text completion task. Machine-learned model(s) 1 can be configured to process input(s) 2 that represent textual data and to generate output(s) 3 that represent additional textual data that completes a textual sequence that includes input(s) 2. For instance, machine-learned model(s) 1 can be configured to generate output(s) 3 to complete a sentence, paragraph, or portion of text that follows from a portion of text represented by input(s) 2.

[0299] In some implementations, the task can be an instruction-following task. Machine-learned model(s) 1 can be configured to process input(s) 2 that represent instructions to perform a function and to generate output(s) 3 that advance a goal of satisfying the instruction function (e.g., at least a step of a multi-step procedure to perform the function). Output(s) 3 can represent data of the same or of a different modality as input(s) 2. For instance, input(s) 2 can represent textual data (e.g., natural language instructions for a task to be performed) and machine-learned model(s) 1 can process input(s) 2 to generate output(s) 3 that represent textual data responsive to the instructions (e.g., natural language responses, programming language responses, machine language responses, etc.). Input(s) 2 can represent image data (e.g., image-based instructions for a task to be performed, optionally accompanied by textual instructions) and machine-learned model(s) 1 can process input(s) 2 to generate output(s) 3 that represent textual data responsive to the instructions (e.g., natural language responses, programming language responses, machine language responses, etc.). One or more output(s) 3 can be iteratively or recursively generated to sequentially process and accomplish steps toward accomplishing the requested functionality. For instance, an initial output can be executed by an external system or be processed by machine-learned model(s) 1 to complete an initial step of performing a function. Multiple steps can be performed, with a final output being obtained that is responsive to the initial instructions.

[0300] In some implementations, the task can be a question answering task. Machine-learned model(s) 1 can be configured to process input(s) 2 that represent a question to answer and to generate output(s) 3 that advance a goal of returning an answer to the question (e.g., at least a step of a multi-step procedure to perform the function). Output(s) 3 can represent data of the same or of a different modality as input(s) 2. For instance, input(s) 2 can represent textual data (e.g., natural language instructions for a task to be performed) and machine-learnedmodel(s) 1 can process input(s) 2 to generate output(s) 3 that represent textual data responsive to the question (e.g., natural language responses, programming language responses, machine language responses, etc.). Input(s) 2 can represent image data (e.g., image-based instructions for a task to be performed, optionally accompanied by textual instructions) and machine-learned model(s) 1 can process input(s) 2 to generate output(s) 3 that represent textual data responsive to the question (e.g., natural language responses, programming language responses, machine language responses, etc.). One or more output(s) 3 can be iteratively or recursively generated to sequentially process and accomplish steps toward answering the question. For instance, an initial output can be executed by an external system or be processed by machine-learned model(s) 1 to complete an initial step of obtaining an answer to the question (e.g., querying a database, performing a computation, executing a script, etc.). Multiple steps can be performed, with a final output being obtained that is responsive to the question.

[0301] In some implementations, the task can be an image generation task. Machine-learned model(s) 1 can be configured to process input(s) 2 that represent context regarding a desired portion of image content. The context can include text data, image data, audio data, etc. Machine-learned model(s) 1 can be configured to generate output(s) 3 that represent image data that depicts imagery related to the context. For instance, machine-learned model(s) 1 can be configured to generate pixel data of an image. Values for channel(s) associated with the pixels in the pixel data can be selected based on the context (e.g., based on a probability determined based on the context).

[0302] In some implementations, the task can be an audio generation task. Machine-learned model(s) 1 can be configured to process input(s) 2 that represent context regarding a desired portion of audio content. The context can include text data, image data, audio data, etc. Machine-learned model(s) 1 can be configured to generate output(s) 3 that represent audio data related to the context. For instance, machine-learned model(s) 1 can be configured to generate waveform data in the form of an image (e.g., a spectrogram). Values for channel(s) associated with pixels of the image can be selected based on the context. Machine-learned model(s) 1 can be configured to generate waveform data in the form of a sequence of discrete samples of a continuous waveform. Values of the sequence can be selected based on the context (e.g., based on a probability determined based on the context).

[0303] In some implementations, the task can be a data generation task. Machine-learned model(s) 1 can be configured to process input(s) 2 that represent context regarding a desired portion of data (e.g., data from various data domains, such as sensor data, image data, multimodal data, statistical data, etc.). The desired data can be, for instance, synthetic data for training other machine-learned models. The context can include arbitrary data type(s). Machine-learned model(s) 1 can be configured to generate output(s) 3 that represent data that aligns with the desired data. For instance, machine-learned model(s) 1 can be configured to generate data values for populating a dataset. Values for the data object(s) can be selected based on the context (e.g., based on a probability determined based on the context).

[0304] Figure 26 is a block diagram of an example networked computing system that can perform aspects of example implementations of the present disclosure. The system can include a number of computing devices and systems that are communicatively coupled over a network 49. An example computing device 50 is described to provide an example of a computing device that can perform any aspect of the present disclosure (e.g., implementing model host 31, client(s) 32, or both). An example server computing system 60 is described as an example of a server computing system that can perform any aspect of the present disclosure (e.g., implementing model host 31, client(s) 32, or both). Computing device 50 and server computing system(s) 60 can cooperatively interact (e.g., over network 49) to perform any aspect of the present disclosure (e.g., implementing model host 31, client(s) 32, or both). Model development platform system 70 is an example system that can host or serve model development platform(s) 12 for development of machine-learned models. Third-party system(s) 80 are example system(s) with which any of computing device 50, server computing system(s) 60, or model development platform system(s) 70 can interact in the performance of various aspects of the present disclosure (e.g., engaging third-party tools, accessing third-party databases or other resources, etc ).

[0305] Network 49 can be any type of communications network, such as a local area network (e.g., intranet), wide area network (e.g., Internet), or some combination thereof and can include any number of wired or wireless links. In general, communication over network 49 can be carried via any type of wired or wireless connection, using a wide variety of communication protocols (e.g., TCP / IP, HTTP, SMTP, FTP), encodings or formats (e.g., HTML, XML), or protection schemes (e.g., VPN, secure HTTP, SSL). Network 49 can also be implemented via asystem bus. For instance, one or more devices or systems of Figure 26 can be co-located with, contained by, or otherwise integrated into one or more other devices or systems.

[0306] Computing device 50 can be any type of computing device, such as, for example, a personal computing device (e.g., laptop or desktop), a mobile computing device (e.g., smartphone or tablet), a gaming console or controller, a wearable computing device, an embedded computing device, a server computing device, a virtual machine operating on a host device, or any other type of computing device. Computing device 50 can be a client computing device. Computing device 50 can be an end-user computing device. Computing device 50 can be a computing device of a service provided that provides a service to an end user (who may use another computing device to interact with computing device 50).

[0307] Computing device 50 can include one or more processors 51 and a memory 52. Processor(s) 51 can be any suitable processing device (e.g., a processor core, a microprocessor, an ASIC, an FPGA, a controller, a microcontroller, etc.) and can be one processor or a plurality of processors that are operatively connected. Memory 52 can include one or more non-transitory computer-readable storage media, such as HBM, RAM, ROM, EEPROM, EPROM, flash memory devices, magnetic disks, etc., and combinations thereof. Memory 52 can store data 53 and instructions 54 which can be executed by processor(s) 51 to cause computing device 50 to perform operations. The operations can implement any one or multiple features described herein. The operations can implement example methods and techniques described herein.

[0308] Computing device 50 can also include one or more input components that receive user input. For example, a user input component can be a touch-sensitive component (e.g., a touch-sensitive display screen or a touch pad) that is sensitive to the touch of a user input object (e.g., a finger or a stylus). The touch-sensitive component can serve to implement a virtual keyboard. Other example user input components include a microphone, camera, LIDAR, a physical keyboard or other buttons, or other means by which a user can provide user input.

[0309] Computing device 50 can store or include one or more machine-learned models 55. Machine-learned models 55 can include one or more machine-learned model(s) 1, such as a sequence processing model 4. Machine-learned models 55 can include one or multiple model instance(s) 31-1. Machine-learned model(s) 55 can be received from server computing system(s) 60, model development platform system 70, third party system(s) 80 (e.g., an application distribution platform), or developed locally on computing device 50. Machine-learned model(s)55 can be loaded into memory 52 and used or otherwise implemented by processor(s) 51.Computing device 50 can implement multiple parallel instances of machine-learned model(s) 55.

[0310] Server computing system (s) 60 can include one or more processors 61 and a memory 62. Processor(s) 61 can be any suitable processing device (e.g., a processor core, a microprocessor, an ASIC, an FPGA, a controller, a microcontroller, etc.) and can be one processor or a plurality of processors that are operatively connected. Memory 62 can include one or more non-transitory computer-readable storage media, such as HBM, RAM, ROM, EEPROM, EPROM, flash memory devices, magnetic disks, etc., and combinations thereof. Memory 62 can store data 63 and instructions 64 which can be executed by processor(s) 61 to cause server computing system(s) 60 to perform operations. The operations can implement any one or multiple features described herein. The operations can implement example methods and techniques described herein.

[0311] In some implementations, server computing system 60 includes or is otherwise implemented by one or multiple server computing devices. In instances in which server computing system 60 includes multiple server computing devices, such server computing devices can operate according to sequential computing architectures, parallel computing architectures, or some combination thereof.

[0312] Server computing system 60 can store or otherwise include one or more machine-learned models 65. Machine-learned model(s) 65 can be the same as or different from machine-learned model(s) 55. Machine-learned models 65 can include one or more machine-learned model(s) 1, such as a sequence processing model 4. Machine-learned models 65 can include one or multiple model instance(s) 31-1. Machine-learned model(s) 65 can be received from computing device 50, model development platform system 70, third party system(s) 80, or developed locally on server computing system(s) 60. Machine-learned model(s) 65 can be loaded into memory 62 and used or otherwise implemented by processor(s) 61. Server computing system(s) 60 can implement multiple parallel instances of machine-learned model(s) 65.

[0313] In an example configuration, machine-learned models 65 can be included in or otherwise stored and implemented by server computing system 60 to establish a client-server relationship with computing device 50 for serving model inferences. For instance, server computing system(s) 60 can implement model host 31 on behalf of client(s) 32 on computing device 50. For instance, machine-learned models 65 can be implemented by server computingsystem 60 as a portion of a web service (e.g., remote machine-learned model hosting service, such as an online interface for performing machine-learned model operations over a network on server computing system(s) 60). For instance, server computing system(s) 60 can communicate with computing device 50 over a local intranet or internet connection. For instance, computing device 50 can be a workstation or endpoint in communication with server computing system(s) 60, with implementation of machine-learned models 65 being managed by server computing system(s) 60 to remotely perform inference (e.g., for runtime or training operations), with output(s) returned (e.g., cast, streamed, etc.) to computing device 50. Machine-learned models 65 can work cooperatively or interoperatively with machine-learned models 55 on computing device 50 to perform various tasks.

[0314] Model development platform system(s) 70 can include one or more processors 71 and a memory 72. Processor(s) 71 can be any suitable processing device (e.g., a processor core, a microprocessor, an ASIC, an FPGA, a controller, a microcontroller, etc.) and can be one processor or a plurality of processors that are operatively connected. Memory 72 can include one or more non-transitory computer-readable storage media, such as HBM, RAM, ROM, EEPROM, EPROM, flash memory devices, magnetic disks, etc., and combinations thereof. Memory 72 can store data 73 and instructions 74 which can be executed by processor(s) 71 to cause model development platform system(s) 70 to perform operations. The operations can implement any one or multiple features described herein. The operations can implement example methods and techniques described herein. Example operations include the functionality described herein with respect to model development platform 12. This and other functionality can be implemented by developer tool(s) 75.

[0315] Third-party system(s) 80 can include one or more processors 81 and a memory 82. Processor(s) 81 can be any suitable processing device (e.g., a processor core, a microprocessor, an ASIC, an FPGA, a controller, a microcontroller, etc.) and can be one processor or a plurality of processors that are operatively connected. Memory 82 can include one or more non-transitory computer-readable storage media, such as HBM, RAM, ROM, EEPROM, EPROM, flash memory devices, magnetic disks, etc., and combinations thereof. Memory 82 can store data 83 and instructions 84 which can be executed by processor(s) 81 to cause third-party system(s) 80 to perform operations. The operations can implement any one or multiple features described herein. The operations can implement example methods and techniques describedherein. Example operations include the functionality described herein with respect to tools and other external resources called when training or performing inference with machine-learned model(s) 1, 4, 16, 20, 55, 65, etc. (e.g., third-party resource(s) 85).

[0316] Figure 26illustrates one example arrangement of computing systems that can be used to implement the present disclosure. Other computing system configurations can be used as well. For example, in some implementations, one or both of computing system 50 or server computing system(s) 60 can implement all or a portion of the operations of model development platform system 70. For example, computing system 50 or server computing system(s) 60 can implement developer tool(s) 75 (or extensions thereof) to develop, update / train, or refine machine-learned models 1, 4, 16, 20, 55, 65, etc. using one or more techniques described herein with respect to model alignment toolkit 17. In this manner, for instance, computing system 50 or server computing system(s) 60 can develop, update / train, or refine machine-learned models based on local datasets (e.g., for model personalization / customization, as permitted by user data preference selections).

[0317] Figure 27 is a block diagram of an example computing device 98 that performs according to example embodiments of the present disclosure. Computing device 98 can be a user computing device or a server computing device (e.g., computing device 50, server computing system(s) 60, etc.). Computing device 98 can implement model host 31. For instance, computing device 98 can include a number of applications (e.g., applications 1 through N). Each application can contain its own machine learning library and machine-learned model(s). For example, each application can include a machine-learned model. Example applications include a text messaging application, an email application, a dictation application, a virtual keyboard application, a browser application, etc. As illustrated in Figure 27, each application can communicate with a number of other components of the computing device, such as, for example, one or more sensors, a context manager, a device state component, or additional components. In some implementations, each application can communicate with each device component using an API (e.g., a public API). In some implementations, the API used by each application is specific to that application.

[0318] Figure 28 is a block diagram of an example computing device 99 that performs according to example embodiments of the present disclosure. Computing device 99 can be the same as or different from computing device 98. Computing device 99 can be a user computingdevice or a server computing device (e.g., computing device 50, server computing system(s) 60, etc.). Computing device 98 can implement model host 31. For instance, computing device 99 can include a number of applications (e.g., applications 1 through N). Each application can be in communication with a central intelligence layer. Example applications include a text messaging application, an email application, a dictation application, a virtual keyboard application, a browser application, etc. In some implementations, each application can communicate with the central intelligence layer (and model(s) stored therein) using an API (e.g., a common API across all applications).

[0319] The central intelligence layer can include a number of machine-learned models. For example, as illustrated in Figure 28, a respective machine-learned model can be provided for each application and managed by the central intelligence layer. In other implementations, two or more applications can share a single machine-learned model. For example, in some implementations, the central intelligence layer can provide a single model for all of the applications. In some implementations, the central intelligence layer is included within or otherwise implemented by an operating system of computing device 99.

[0320] The central intelligence layer can communicate with a central device data layer. The central device data layer can be a centralized repository of data for computing device 99. As illustrated in Figure 28, the central device data layer can communicate with a number of other components of the computing device, such as, for example, one or more sensors, a context manager, a device state component, or additional components. In some implementations, the central device data layer can communicate with each device component using an API (e.g., a private API).

[0321] The technology discussed herein makes reference to servers, databases, software applications, and other computer-based systems, as well as actions taken and information sent to and from such systems. The inherent flexibility of computer-based systems allows for a great variety of possible configurations, combinations, and divisions of tasks and functionality between and among components. For instance, processes discussed herein can be implemented using a single device or component or multiple devices or components working in combination. Databases and applications can be implemented on a single system or distributed across multiple systems. Distributed components can operate sequentially or in parallel.

[0322] While the present subject matter has been described in detail with respect to various specific example embodiments thereof, each example is provided by way of explanation, not limitation of the disclosure. Those skilled in the art, upon attaining an understanding of the foregoing, can readily produce alterations to, variations of, and equivalents to such embodiments. Accordingly, the subject disclosure does not preclude inclusion of such modifications, variations or additions to the present subject matter as would be readily apparent to one of ordinary skill in the art. For instance, features illustrated or described as part of one embodiment can be used with another embodiment to yield a still further embodiment. Thus, it is intended that the present disclosure cover such alterations, variations, and equivalents.

[0323] Aspects of the disclosure have been described in terms of illustrative embodiments thereof. Any and all features in the following claims can be combined or rearranged in any way possible, including combinations of claims not explicitly enumerated in combination together, as the example claim dependencies listed herein should not be read as limiting the scope of possible combinations of features disclosed herein. Accordingly, the scope of the present disclosure is by way of example rather than by way of limitation, and the subject disclosure does not preclude inclusion of such modifications, variations or additions to the present subject matter as would be readily apparent to one of ordinary skill in the art. Moreover, terms are described herein using lists of example elements joined by conjunctions such as “and,” “or,” “but,” etc. It should be understood that such conjunctions are provided for explanatory purposes only. Clauses and other sequences of items joined by a particular conjunction such as “or,” for example, can refer to “and / or,” “at least one of’, “any combination of’ example elements listed therein, etc. Terms such as “based on” should be understood as “based at least in part on.”

[0324] The term “can” should be understood as referring to a possibility of a feature in various implementations and not as prescribing an ability that is necessarily present in every implementation. For example, the phrase “X can perform Y” should be understood as indicating that, in various implementations, X has the potential to be configured to perform Y, and not as indicating that in every instance X must always be able to perform Y. It should be understood that, in various implementations, X might be unable to perform Y and remain within the scope of the present disclosure.

[0325] The term “may” should be understood as referring to a possibility of a feature in various implementations and not as prescribing an ability that is necessarily present in every implementation. For example, the phrase “X may perform Y” should be understood as indicating that, in various implementations, X has the potential to be configured to perform Y, and not as indicating that in every instance X must always be able to perform Y. It should be understood that, in various implementations, X might be unable to perform Y and remain within the scope of the present disclosure.

Claims

WHAT IS CLAIMED IS:

1. A method implemented by one or more computing devices, comprising:receiving voice input that conveys a user instruction with reference to a displayed user interface;monitoring a location of a point of attention of the user on the displayed user interface;identifying one or more elements of the displayed user interface based on the location of the point of attention at one or more times;forming a prompt for a generative model based on the voice input and the one or more elements;obtaining computer code, comprising sending the prompt to the generative model, andexecuting the computer code.

2. The method of claim 1, comprising evaluating the voice input for one or more trigger events, wherein identifying the one or more elements of the displayed user interface comprises temporally correlating the location of the point of attention with the one or more trigger events.

3. The method of claim 2, wherein the one or more trigger events comprise the detection of one or more words in the voice input.

4. The method of any one of the preceding claims, wherein forming the prompt comprises generating an augmented transcription of at least part of the voice input and including the augmented transcription in the prompt, wherein the augmented transcription includes data relating to the one or more elements.

5. The method of claim 4, wherein the data relating to the one or more elements comprises structured data for at least one of the one or more elements.

6. The method of claim 4 or claim 5, wherein the data relating to the one or more elements comprises image data for at least one of the one or more elements.

7. The method of any one of claims 4 to 6, wherein the data relating to the one or more elements comprises a reference to at least one of the one or more elements.

8. The method of any one of the preceding claims, wherein the point of attention is identified by the position of a cursor on the displayed user interface.

9. The method of claim 8, wherein the cursor comprises a mouse pointer.

10. The method of any one of the preceding claims, wherein the point of attention is identified by a touch input on the displayed user interface.

11. The method of any one of the preceding claims, wherein forming the prompt comprises including, in the prompt, structured data for at least one of the one or more elements.

12. The method of any one of the preceding claims, wherein forming the prompt comprises including, in the prompt, a reference to at least one of the one or more elements.

13. The method of any one of the preceding claims, wherein forming the prompt further comprises including, in the prompt, content relating to at least one of the one or more elements.

14. The method of any one of the preceding claims, further comprising:receiving a signal generated responsive to selection of a hardware selector configured for initiating an enhanced interaction mode for interacting with the displayed user interface, and initiating the enhanced interaction mode responsive to receiving the signal, wherein the voice input that conveys the user instruction is received after the enhanced interaction mode has been initiated.

15. The method of claim 14, wherein the hardware selector configured to initiate the enhanced interaction mode is a physical selector disposed on a mouse or a keyboard or is a virtual selector corresponding to a portion of a touch- sensitive interface.

16. The method of any one of the preceding claims, further comprising:receiving a signal generated responsive to selection of a hardware selector configured for initiating a mode of the displayed user interface in which a plurality of elements of the displayed user interface may be selected with a single trigger event,initiating the mode responsive to receiving the signal,identifying a plurality of elements of the displayed user interface based on the location of the point of attention at or after the single trigger event.

17. The method of claim 16 when dependent on claim 14 or claim 15, wherein the hardware selector configured for initiating the enhanced interaction mode is the same as the hardware selector for initiating the mode of the displayed user interface in which the plurality of elements of the displayed user interface may be selected with a single trigger event.

18. The method of any one of the preceding claims, wherein at least one of the one or more elements comprises image data for at least part of the displayed user interface, and wherein forming the prompt comprises including, in the prompt, the image data.

19. The method of any one of claims 1 to 18, wherein:identifying the one or more elements comprises extracting image data for one or more regions of the displayed user interface, each region corresponding to the location of the point of attention at one or more of the one or more times, andforming the prompt, comprising including the image data for the one or more regions in the prompt.

20. The method of any one of the preceding claims, wherein the one or more elements include an element relating to a first resource, and wherein the user instruction comprises an instruction to transfer data specified by the element to a second resource.

21. The method of claim 20, wherein the one or more elements further comprise a second element relating to the second resource.

22. The method of any one of the preceding claims, wherein the user instruction comprises an instruction to fill out a form based on at least one of the one or more elements.

23. The method of any one of the preceding claims, wherein the user instruction comprises an instruction to alter one or more attributes of at least one of the one or more elements.

24. The method of any one of the preceding claims, wherein the user instruction comprises an instruction to conditionally link at least two of the elements.

25. The method of any one of the preceding claims, wherein the user instruction comprises an instruction to modify an image.

26. The method of any one of the preceding claims, wherein the user instruction comprises an instruction to edit a document, spreadsheet or chart.

27. The method of any one of the preceding claims, wherein the user instruction comprises an instruction to add comments or logging functionality to computer code.

28. The method of any one of the preceding claims, wherein executing the code causes the user instruction to be carried out.

29. The method of any one of the preceding claims, further comprising highlighting the one or more elements on the displayed user interface.

30. The method of any one of the preceding claims, further comprising, responsive to executing the code, providing audio and / or text feedback corresponding to the user instruction.

31. A method implemented by one or more computing devices, comprising:obtaining a context stream comprising a stream of context data over time;receiving user input that refers to the context stream;evaluating the user input for one or more trigger events;determining one or more prompt items by sampling from the context stream in temporal correspondence with the one or more trigger events, andforming a prompt for a generative model using the one or more prompt items.

32. The computer-implemented method of claim 31, wherein the user input comprises voice input, wherein the prompt is formed based on the voice input and the one or more prompt items.

33. The computer-implemented method of claim 32, wherein the one or more trigger events comprise the detection of one or more words in the voice input, wherein evaluating the user input for one or more trigger events comprises analysing the voice input to detect the one or more words.

34. The computer-implemented method of any one of claims 31 to 33, wherein at least one of the one or more trigger events comprise at least one selection of a hardware selector.

35. The computer-implemented method of any one of claims 31-34, wherein forming the prompt comprises inserting the one or more prompt items into text corresponding to a user instruction that is conveyed by the user input.

36. The computer-implemented method of claim 35 when dependent on claim 32, wherein the text comprises a transcription of at least a part of the voice input.

37. The computer-implemented method of any one of claims 31-36, wherein the one or more prompt items comprise structured data for one or more visual components that are displayed on a display.

38. The computer-implemented method of any one of claims 31 -37, wherein the one or more prompt items comprise image data for one or more image regions that are shown on a display.

39. The computer-implemented method of any one of claims 31-38, wherein the context stream comprises locations on a display, the locations comprising cursor locations or touch locations.

40. The computer-implemented method of any one of claims 31-39, further comprising obtaining code based on the prompt, comprising sending the prompt to the generative model.

41. The computer-implemented method of claim 40, further comprising executing the code.

42. A method implemented by one or more computing devices, comprising:monitoring a location of a point of attention of a user on a displayed user interface,determining one or more prompt items, comprising determining at least one of the prompt items based on the location of the point of attention of the user at a respective time, wherein each of the one or more prompt items includes a corresponding timestamp, and forming a prompt for a generative model based on the one or more prompt items.

43. The method of claim 42, comprising receiving a voice input, wherein the one or more prompt items includes at least one prompt item comprising a transcription of at least part of the voice input.

44. The method of claim 42 or claim 43, further comprising obtaining structured data for an element of the displayed user interface based on the location of the point of attention of the user at the respective time, wherein the at least one prompt item which is determined based on the location of the point of attention of the user at the respective time comprises the structured data.

45. The computer-implemented method of claim 44, wherein the structured data comprises one or more markup language elements.

46. The computer-implemented method of any one of claim 44 or claim 45, wherein the structured data comprises a reference to the corresponding element of the displayed user interface.

47. The computer-implemented method of any one of claims 42 to 46, wherein the location of the point of attention of the user on the displayed user interface is specified by a cursor location or a touch input location.

48. The computer-implemented method of any one of claims 42-47, further comprising sending the prompt to the generative model and receiving a corresponding response.

49. One or more non-transitory computer-readable media storing instructions that are executable by one or more processors to cause a computing system to perform the method of any one of the preceding claims.

50. A computing system, comprising:one or more processors; andone or more non-transitory computer-readable media storing instructions that are executable by the one or more processors to cause the computing system to perform operations, the operations comprising the method of any one of the preceding claims.

Images