Cloud-assisted in-vehicle large-scale language model

Abstract

To properly use an LLM (large scale language model) in a vehicle.SOLUTION: Inputting a pre-redacted prompt to a first 1LLM implemented in the vehicle, wherein the pre-redacted prompt includes personalized information associated with an occupant of the vehicle, and transmitting a post-redacted prompt to the first 2LLM implemented in the cloud, wherein the post-redacted prompt represents the pre-redacted prompt without the personalized information; Receiving the logit of the second 2LLM, inputting the edited prompt into the second 3LLM implemented in the vehicle or in the cloud, wherein the second 3LLM includes the same set of parameters as the first 1LLM, determining a difference between the logit of the second 1LLM and the logit of the first 3LLM, determining a sum of the difference and the logit of the first 2LLM to obtain an adjusted logit, and obtaining an output from the adjusted logit.SELECTED DRAWING: Figure 1

Description

[Background technology]

[0001] To improve the user experience in vehicles, large language models (LLMs) are used to perform vehicle control, provide appropriate information to occupants, etc. LLMs can be executed using local computing resources or cloud-based computing resources. [Brief explanation of the drawings]

[0002] Aspects of the present disclosure are best understood by reading the following detailed description in conjunction with the accompanying drawings. It should be noted that, according to industry practice, the various drawings are not drawn to scale. In fact, dimensions in the various drawings may be arbitrarily increased or reduced for clarity of presentation.

[0003] [Figure 1] FIG. 1 is a schematic diagram of a system for cloud-assisted in-vehicle large-scale language models, according to at least some embodiments of the present disclosure.

[0004] [Figure 2] FIG. 2 is a schematic diagram of a system for cloud-assisted in-vehicle large-scale language models, according to at least some embodiments of the present disclosure.

[0005] [Figure 3] FIG. 3 is a schematic diagram of a system for cloud-assisted in-vehicle large-scale language models, according to at least some embodiments of the present disclosure.

[0006] [Figure 4] FIG. 4 is a schematic diagram of a system for cloud-assisted in-vehicle large-scale language models, according to at least some embodiments of the present disclosure.

[0007] [Figure 5] FIG. 5 is an operational flow diagram for cloud-assisted in-vehicle large-scale language models, according to at least some embodiments of the present disclosure.

[0008] [Figure 6] FIG. 6 is an operational flow diagram for preparing a pre-edit prompt in accordance with at least some embodiments of the present disclosure.

[0009] [Figure 7] FIG. 7 is a schematic diagram of a vehicle, according to at least some embodiments of the present disclosure.

[0010] [Figure 8] FIG. 8 is an operational flow diagram for logit adjustment, in accordance with at least some embodiments of the present disclosure.

[0011] [Figure 9] FIG. 9 is an operational flow diagram for prompt editing in accordance with at least some embodiments of the present disclosure.

[0012] [Figure 10] FIG. 10 is a schematic diagram of a hardware configuration for use of cloud-assisted in-vehicle large-scale language models, in accordance with at least some embodiments of the present disclosure. DETAILED DESCRIPTION OF THE INVENTION

[0013] The following disclosure provides numerous different embodiments or examples for implementing different features of the presented subject matter. To simplify the disclosure, certain examples of components, values, operations, materials, combinations, etc. are described below. These are, of course, merely examples and are not intended to be limiting. Other components, values, operations, materials, combinations, etc. are also contemplated. In addition, the disclosure may repeat reference numerals and / or characters in different examples. This repetition is for the purpose of simplicity and clarity and does not in itself indicate a relationship between the various embodiments and / or configurations in which this occurs.

[0014] LLMs implemented in vehicles may have limited accuracy in their output because the size of the LLM may be limited by vehicle capacity and processing constraints. The use of LLMs implemented in the cloud may increase the risk of leakage of personal information, including personal information of passengers, contained in the input to the LLM. Furthermore, because the output from inputs that include personal information may be more accurate than the output from inputs that do not, removing such personal information from the input may limit the accuracy of the output.

[0015] In at least some embodiments of the present disclosure, a more accurate output can be achieved by editing the pre-edit prompt to remove personal information, comparing the logits of the pre-edit and post-edit prompts using one or two smaller LLMs, and applying the difference to the logits of the larger LLMs to obtain an adjusted logit. In at least some embodiments, the logit is an intermediate output of the LLM, represented in the form of a non-normalized vector containing values ​​representing confidence in the most appropriate candidate for output. For example, if the prompt input to the LLM is "What material is this road made of?", the logit may include values ​​representing a 95% confidence in "asphalt," a 94.6% confidence in "concrete," a 91.2% confidence in "dirt," and an 84.3% confidence in "gravel." In at least some embodiments, the difference in the logits represents the mathematical difference between the values ​​representing the confidence in each of the common candidates. For example, if the logit of the smaller LLM from the pre-edit prompt contains a value representing a 95% confidence in “asphalt,” and the logit of the smaller LLM from the post-edit prompt contains a value representing a 94.7% confidence in “asphalt,” the difference in values ​​represents approximately a 0.3% increase in the confidence in “asphalt.” Note that converting the unnormalized logit value to a confidence percentage does not capture this exact difference in confidence. In at least some embodiments, this increase in confidence is presumed to be due to the inclusion of personal information in the input. In at least some embodiments, applying the difference to the logit represents a mathematical sum of the difference value and the logit value representing the candidate's confidence. For example, if the logit of the larger LLM from the post-edit prompt contains a value representing a 94.2% confidence in “asphalt,” applying the difference value representing a 0.3% increase results in an adjusted logit value representing a 94.5% confidence in “asphalt.” Note that converting unnormalized logit values ​​to percent confidence does not result in an exact sum of such confidences. In at least some embodiments, the most likely candidate according to the adjusted logit may differ from the most likely candidate according to the logit of the smaller LLM from the pre-edit prompt and the most likely candidate according to the logit of the larger LLM from the post-edit prompt.

[0016] In at least some embodiments, using a large LLM in the cloud without sending personal information to the cloud allows for more accurate output from the vehicle's LLM, and in at least some embodiments, using other smaller LLMs in the cloud allows for a lower computational load on the vehicle.

[0017] In at least some embodiments, all LLMs have logits of the same dimension. In at least some embodiments, one or two smaller LLMs have the same parameter set with different parameter values. In at least some embodiments, one of the two smaller LLMs is trained on personal information, and the other of the two smaller LLMs and the larger LLM are trained generically. In at least some embodiments, the difference between the logits is calculated from the logits from the pre-edit prompts and the logits from the post-edit prompts output by one of the smaller LLMs. In at least some embodiments, using one smaller LLM to calculate the difference can reduce resource consumption.

[0018] In at least some embodiments, the vehicle has at least one small LLM and the large LLM is in the cloud. In at least some embodiments, other small LLMs may be included in either the vehicle or the cloud. In at least some embodiments, a logit comparison between the outputs of at least one small LLM corresponding to the pre-edit prompt and the post-edit prompt is performed in the vehicle or the cloud. In at least some embodiments, if other commonly learned small LLMs are in the cloud, performing the comparison in the cloud makes sense. In at least some embodiments, the difference and the logits of the large LLM are summed in the vehicle or the cloud. In at least some embodiments, the iterative outputs are summed in the vehicle or the cloud.

[0019] In at least some embodiments, the raw prompt may be an instruction, a query, a command, or the like. In at least some embodiments, the raw prompt may be combined with an edited prompt to provide context for creating the pre-edited prompt. For example, the edited prompt may add predetermined information, such as information that the user is driving a vehicle. In at least some embodiments, the raw prompt may be enriched with personal information relevant to the raw prompt, such as a precise location, allergy information, recent activities, known preferences, and the like, to create the pre-edited prompt. In at least some embodiments, Retrieval Augmented Generation (RAG) techniques may be used to determine the relevant personal information. Examples of personal information, such as passenger information, include facial images of the passenger (e.g., the driver), voice recordings of the passenger, identification information (e.g., MAC address) of the passenger's mobile device, vehicle operation by the passenger, information about the passenger previously stored in a storage device, and the like. In at least some embodiments, the pre-edited prompt is edited into a post-edited prompt by, for example, text-to-speech conversion, image data to object data conversion, location generalization, filtering the raw prompt, randomizing the raw prompt, and the like.

[0020] 1 is a schematic diagram of a system for cloud-assisted in-vehicle large-scale language models according to at least some embodiments of the present disclosure. The system includes a vehicle 100, a cloud network 109, a first LLM 110, pre-edit prompts 120, personal information 112, a first LLM logit 121, a second LLM 114, post-edit prompts 123, a second LLM logit 124, a third LLM 116, a third LLM logit 126, a logit difference 127, and a logit sum 128.

[0021] Vehicle 100 includes first LLM 110 and personal information 112 and is in communication with cloud network 109. In at least some embodiments, vehicle 100 is configured to support installation and operation of first LLM 110. In at least some embodiments, vehicle 100 is configured to provide transportation to its occupants. In at least some embodiments, vehicle 100 is configured to input pre-edit prompts 120 into first LLM 110. In at least some embodiments, vehicle 100 is configured to receive third LLM logits 126 from cloud network 109. In at least some embodiments, vehicle 100 is configured to calculate logit difference 127 by comparing first LLM logit 121 and third LLM logit 126. In at least some embodiments, vehicle 100 is configured to receive second LLM logits 124 from cloud network 109. In at least some embodiments, vehicle 100 is configured to calculate logit sum 128 by adding logit difference 127 to second LLM logit 124. In at least some embodiments, vehicle 100 includes a network, such as a Controller Area Network (CAN), configured to transmit personal information 112 to first LLM 110. In at least some embodiments, vehicle 100 is one of many vehicles equipped with an on-board computing system. In at least some embodiments, vehicle 100 is an automobile, truck, boat, aircraft, submarine, etc. In at least some embodiments, vehicle 100 is configured for transportation, navigation, recreation, etc.

[0022] Cloud network 109 is in communication with vehicle 100 and includes second LLM 114 and third LLM 116. In at least some embodiments, cloud network 109 is configured to transmit data between vehicle 100 and second LLM 114 and third LLM 116. In at least some embodiments, cloud network 109 is configured to receive post-edit prompt 123 from vehicle 100 and input post-edit prompt 123 into second LLM 114 and third LLM 116. In at least some embodiments, cloud network 109 is configured to transmit second LLM logits 124 and third LLM logits 126 to vehicle 100. In at least some embodiments, cloud network 109 is configured to provide a platform for various cloud-based services. In at least some embodiments, cloud network 109 is configured to interact with internet service providers and data centers. In at least some embodiments, cloud network 109 includes a data center having servers that host cloud services. In at least some embodiments, cloud network 109 is configured for data storage, web hosting, Software as a Service (SaaS), and the like.

[0023] The first LLM 110 is included in the vehicle 100 and is configured to process prompts and generate logits. In at least some embodiments, the first LLM 110 is configured to process pre-edited prompts 120 and generate first LLM logits 121. In at least some embodiments, the first LLM 110 is configured to interact with personal information 112. In at least some embodiments, the first LLM 110 has been trained for various language model tasks, such as natural language processing, speech recognition, image processing, etc. In at least some embodiments, the first LLM 110 has been trained, at least in part, using personal information 1120. In at least some embodiments, the first LLM 110 is configured to interact with other components of the vehicle 100. In at least some embodiments, the first LLM 110 includes a software module installed on a computing system of the vehicle 100. In at least some embodiments, the first LLM 110 includes a neural network, instructions for processing input through the neural network, and instructions for converting logits into output that can be interpreted by a human or a computer.

[0024] Pre-edit prompt 120 provides the initial input to first LLM 110 and serves as the basis for post-edit prompt 123. In at least some embodiments, pre-edit prompt 120 includes personal information, such as information in personal information 112. In at least some embodiments, pre-edit prompt 120 is used as the basis for creating post-edit prompt 123. In at least some embodiments, pre-edit prompt 120 is a voice command, text input, etc. In at least some embodiments, pre-edit prompt 120 includes a query, command, etc. from a user.

[0025] Personal information 112 is a collection of personal information, such as personal information included in pre-edit prompt 120. In at least some embodiments, personal information 112 is used to edit the personal information included in pre-edit prompt 120 to create post-edit prompt 123. In at least some embodiments, personal information 112 is specific to one or more vehicle occupants. In at least some embodiments, personal information 112 is used to populate live prompts. In at least some embodiments, personal information 112 includes a database that stores and indexes a collection of personal information. In at least some embodiments, the collection of personal information includes a user's name, preferences, etc.

[0026] The first LLM logits 121 are the output of the first LLM 110 based on the pre-edit prompts 120 and are used to calculate the logit difference 127. In at least some embodiments, the first LLM logits 121 represent an intermediate output of the first LLM 110. In at least some embodiments, the first LLM logits 121 are a vector of probabilities for each possible next word in a word sequence. In at least some embodiments, the first LLM logits 121 represent the confidence of each of the most likely candidates for the output.

[0027] The second LLM 114 resides in the cloud network 109 and is configured to process prompts and generate logits. In at least some embodiments, the second LLM 114 is configured to process edited prompts 123 and generate second LLM logits 124. In at least some embodiments, the second LLM 114 has a larger set of parameters than the first LLM 110. In at least some embodiments, the second LLM 114 has more layers than the first LLM 110. In at least some embodiments, the second LLM 114 has more interconnections than the first LLM 110. In at least some embodiments, the second LLM 114 is trained for various language model tasks, such as natural language processing, speech recognition, image processing, etc. In at least some embodiments, the second LLM 114 is trained for general use without personal information 112. In at least some embodiments, the second LLM 114 is configured to interact with vehicles and computers different from the vehicle 100. In at least some embodiments, the second LLM 114 includes a software module that executes using the computational resources of the cloud network 109. In at least some embodiments, the second LLM 114 includes a neural network, instructions for processing inputs through the neural network, and instructions for converting logits into human- or computer-interpretable outputs.

[0028] The post-edit prompt 123 represents the pre-edit prompt 120 with personal information removed. In at least some embodiments, the post-edit prompt 123 is input to the second LLM 114 and the third LLM 116. In at least some embodiments, the post-edit prompt 123 is processed from the pre-edit prompt 120, such as text-to-speech, image data to object data conversion, location generalization, filtering, randomization, etc.

[0029] The second LLM logit 124 is the output of the second LLM 114 based on the post-edit prompt 123 and is used to calculate the logit sum 128. In at least some embodiments, the second LLM logit 124 represents an intermediate output of the second LLM 114. In at least some embodiments, the second LLM logit 124 is a vector of probabilities for each possible next word in a word sequence. In at least some embodiments, the second LLM logit 124 represents the confidence of each of the most likely candidates for the output. In at least some embodiments, the dimensionality of the second LLM logit 124 is the same as that of the first LLM logit 121.

[0030] The third LLM 116 resides in the cloud network 109 and is configured to process prompts and generate logits. In at least some embodiments, the third LLM 116 is configured to process the edited prompts 123 and generate third LLM logits 126. In at least some embodiments, the third LLM 116 has the same set of parameters as the first LLM 110. In at least some embodiments, the third LLM 116 has different parameter values ​​than the first LLM 110. In at least some embodiments, the third LLM 116 is trained for various language model tasks, such as natural language processing, speech recognition, image processing, etc. In at least some embodiments, the third LLM 116 is trained for general use without using personal information 112. In at least some embodiments, the third LLM 116 is configured to interact with vehicles and computers other than the vehicle 100. In at least some embodiments, the third LLM 116 includes software modules that execute using computational resources of the cloud network 109. In at least some embodiments, the third LLM 116 includes a neural network, instructions for processing inputs through the neural network, and instructions for converting logits into human- or computer-interpretable output.

[0031] The third LLM logit 126 is the output of the third LLM 116 based on the post-edit prompt 123 and is used to calculate the logit difference 127. In at least some embodiments, the third LLM logit 126 represents an intermediate output of the third LLM 116. In at least some embodiments, the third LLM logit 126 is a vector of probabilities for each possible next word in a word sequence. In at least some embodiments, the third LLM logit 126 represents the confidence of each of the most likely candidates for the output. In at least some embodiments, the dimensionality of the third LLM logit 126 is the same as the first LLM logit 121 and the second LLM logit 124.

[0032] Logit Difference 127 represents the difference between the first LLM logit 121 and the third LLM logit 126. In at least some embodiments, Logit Difference 127 is used to calculate Logit Sum 128. In at least some embodiments, Logit Difference 127 is a vector of differences in the likelihood of each word being the next word in a word sequence. In at least some embodiments, Logit Difference 127 represents the difference in the confidence of each of the most likely candidates for the output. In at least some embodiments, Logit Difference 127 has the same dimension as the first LLM logit 121, the second LLM logit 124, and the third LLM logit 126.

[0033] Logit sum 128 represents the sum of logit difference 127 and second LLM logit 124. In at least some embodiments, logit sum 128 represents the adjusted logit. In at least some embodiments, logit sum 128 is the adjusted logit used to derive the output. In at least some embodiments, logit sum 128 is a vector of likelihoods for each word that can be the next word in a word sequence. In at least some embodiments, logit sum 128 represents the confidence of each of the most likely candidates for the output. In at least some embodiments, logit sum 128 has the same dimension as first LLM logit 121, second LLM logit 124, and third LLM logit 126.

[0034] 2 is a schematic diagram of a system for cloud-assisted in-vehicle large-scale language models according to at least some embodiments of the present disclosure. The system includes a vehicle 200, a cloud network 209, a first LLM 210, pre-edit prompts 220, personal information 212, a first LLM logit 221, a second LLM 214, a post-edit prompt 223, a second LLM logit 224, a third LLM 216, a third LLM logit 226, a logit difference 227, and a logit sum 228.

[0035] Vehicle 200, cloud network 209, first LLM 210, pre-edit prompt 220, personal information 212, first LLM logit 221, second LLM 214, post-edit prompt 223, second LLM logit 224, third LLM 216, third LLM logit 226, logit difference 227, and logit sum 228 are substantially similar in structure and function to vehicle 100, cloud network 109, first LLM 110, pre-edit prompt 120, personal information 112, first LLM logit 121, second LLM 114, post-edit prompt 123, second LLM logit 124, third LLM 116, third LLM logit 126, logit difference 127, and logit sum 128, respectively, unless otherwise specified.

[0036] 2 , third LLM 216 is included in vehicle 200. In at least some embodiments, third LLM 216 comprises a software module installed on a computing system of vehicle 200. In at least some embodiments, vehicle 200 is configured to receive second LLM logits 224 from cloud network 209, but may not be configured to receive third LLM logits 226 from cloud network 209. In at least some embodiments, vehicle 200 requires more computational resources than vehicle 100 of FIG. 1 , but vehicle 200 requires less communication with cloud network 209 than vehicle 100 requires with cloud network 109 of FIG. 1 .

[0037] 3 is a schematic diagram of a system for cloud-assisted in-vehicle large-scale language models according to at least some embodiments of the present disclosure. The system includes a vehicle 300, a cloud network 309, a first LLM 310, pre-edit prompts 320, personal information 312, first LLM logits 321, a second LLM 314, post-edit prompts 323, second LLM logits 324, a third LLM 316, third LLM logits 326, logit difference 327, and logit sum 328.

[0038] Vehicle 300, cloud network 309, first LLM 310, pre-edit prompt 320, personal information 312, first LLM logit 321, second LLM 314, post-edit prompt 323, second LLM logit 324, third LLM 316, third LLM logit 326, logit difference 327, and logit sum 328 are substantially similar in structure and function to vehicle 100, cloud network 109, first LLM 110, pre-edit prompt 120, personal information 112, first LLM logit 121, second LLM 114, post-edit prompt 123, second LLM logit 124, third LLM 116, third LLM logit 126, logit difference 127, and logit sum 128, respectively, unless otherwise specified.

[0039] 3 , logit difference 327 and logit sum 328 are calculated by cloud network 309. In at least some embodiments, vehicle 300 is configured to transmit first LLM logit 321 to cloud network 309. In at least some embodiments, cloud network 309 is configured to receive first LLM logit 321 from vehicle 300. In at least some embodiments, cloud network 309 is configured to calculate logit difference 327 by comparing first LLM logit 321 with third LLM logit 326. In at least some embodiments, cloud network 309 is configured to calculate logit sum 328 by adding logit difference 327 to second LLM logit 324. In at least some embodiments, cloud network 309 is configured to transmit logit sum 328, or an output calculated therefrom, to vehicle 300. In at least some embodiments, vehicle 300 requires fewer computing resources than vehicle 100 of FIG. 1, and vehicle 300 requires less communication with cloud network 309 than vehicle 200 requires with cloud network 209 of FIG. 2.

[0040] 4 is a schematic diagram of a system for cloud-assisted in-vehicle large-scale language models according to at least some embodiments of the present disclosure. The system includes a vehicle 400, a cloud network 409, a first LLM 410, pre-edit prompts 420, personal information 412, first LLM logits 421, a second LLM 414, post-edit prompts 423, second LLM logits 424, a third LLM 416, third LLM logits 426, logit difference 427, and logit sum 428.

[0041] Vehicle 400, cloud network 409, first LLM 410, pre-edit prompt 420, personal information 412, first LLM logit 421, second LLM 414, post-edit prompt 423, second LLM logit 424, third LLM 416, third LLM logit 426, logit difference 427, and logit sum 428 are substantially similar in structure and function to vehicle 200, cloud network 209, first LLM 210, pre-edit prompt 220, personal information 212, first LLM logit 221, second LLM 214, post-edit prompt 223, second LLM logit 224, third LLM 216, third LLM logit 226, logit difference 227, and logit sum 228, respectively, unless otherwise specified.

[0042] In the system of FIG. 4 , logit sum 428 is calculated by cloud network 409. In at least some embodiments, vehicle 400 is configured to transmit logit difference 427 to cloud network 409. In at least some embodiments, cloud network 409 is configured to receive logit difference 427 from vehicle 400. In at least some embodiments, cloud network 409 is configured to calculate logit sum 428 by adding logit difference 427 to second LLM logit 424. In at least some embodiments, cloud network 409 is configured to transmit logit sum 428, or an output calculated therefrom, to vehicle 400. In at least some embodiments, vehicle 400 requires fewer computational resources than vehicle 100 of FIG. 1 but more than vehicle 400 of FIG. 3 . In at least one embodiment, vehicle 400 requires less communication with cloud network 409 than vehicle 200 requires with cloud network 209 of FIG. 2, but not less than vehicle 300 requires with cloud network 309 of FIG. 3.

[0043] In at least some embodiments other than those of Figures 1-4, the configuration and functionality may be distributed between the vehicle and the cloud network in a different manner such that pre-edit prompts are not sent from the vehicle and the second LLM logits are processed at least in part using the computational resources of the cloud network.

[0044] 5 is an operational flow diagram for a cloud-assisted, in-vehicle large-scale language model, according to at least some embodiments of the present disclosure. In at least some embodiments, the operational flow diagram presents a method for using a cloud-assisted, in-vehicle large-scale language model, according to at least some embodiments of the present disclosure. In at least some embodiments, the method is performed by a controller of a server, such as controller 1002 of server 1000 in FIG. 10, described below.

[0045] At S530, the controller inputs the pre-edit prompt into a first LLM. In at least some embodiments, the controller inputs the pre-edit prompt into a first large-scale language model (LLM) installed in the vehicle. In at least some embodiments, the pre-edit prompt includes personal information. In at least some embodiments, the controller obtains logits output from the first LLM. In at least some embodiments, the logits are used in subsequent operations. In at least some embodiments, the controller executes the operational flow of FIG. 6, described below.

[0046] At S532, the controller adjusts the logits. In at least some embodiments, the controller adjusts the logits output from the first LLM based on other logits of similar LLMs. In at least some embodiments, the controller adjusts the logits based on other logits of similar cloud-based LLMs. In at least some embodiments, the controller calculates the adjusted logits. In at least some embodiments, the controller executes the operational flow of FIG. 8, described below.

[0047] At S533, the controller obtains an iterative output from the adjusted logits. In at least some embodiments, the controller converts the adjusted logits into a human- or computer-interpretable output. In at least some embodiments, the controller obtains the next word in the word string.

[0048] At S534, the controller accumulates the iteration output. In at least some embodiments, the controller accumulates the iteration output from the current iteration with the iteration output of the previous iteration. In at least some embodiments, the controller accumulates the word string following the pre-edit prompt.

[0049] At S536, the controller determines whether the output is complete. If the output is not complete, the operational flow proceeds to S538 to concatenate the iterative output with a pre-edit prompt, and then returns to the prompt input at S530. If the output is complete, the operational flow proceeds to S539 to provide or use the output. In at least some embodiments, the controller determines whether to proceed with further iterations by determining whether the probability of the word string being completed is greater than the probability of other words following.

[0050] At S538, the controller concatenates the iteration output to the pre-edit prompt. In at least some embodiments, the controller concatenates the word string to the pre-edit prompt. In at least some embodiments, the concatenation result is input to the first LLM as an updated pre-edit prompt in the next iteration of S530.

[0051] At S539, the controller provides or uses the output. In at least some embodiments, the controller provides the accumulated output to a vehicle occupant or uses the output to control the vehicle. In at least some embodiments, the controller displays the accumulated output through a vehicle display, plays the accumulated output through a vehicle speaker, etc. In at least some embodiments, the controller uses the accumulated output to control the vehicle or a component thereof, or passes the accumulated output to another component of the vehicle for control of the vehicle or component thereof.

[0052] 6 is an operational flow diagram for preparing a pre-editing prompt, in accordance with at least some embodiments of the present disclosure. In at least some embodiments, the operational flow diagram presents a method for preparing a pre-editing prompt, in accordance with at least some embodiments of the present disclosure. In at least some embodiments, the method is performed by a controller of a server, such as controller 1002 of server 1000 of FIG. 10 , described below.

[0053] At S640, the controller, or a portion thereof, detects a live prompt. In at least some embodiments, the controller identifies and captures the live prompt, such as an audio sample, an image of the occupant, etc. In at least some embodiments, the controller detects the live prompt through a microphone, an image sensor, etc. In at least some embodiments, the controller is operative and waiting to receive the live prompt.

[0054] At S642, the controller, or portions thereof, calculates feature vectors. In at least some embodiments, the controller processes the raw prompts to calculate one or more feature vectors. In at least some embodiments, the controller extracts embedded information and associates it with personal information. In at least some embodiments, the controller transforms the raw prompts into a form that can be analyzed and processed.

[0055] At S644, the controller, or a portion thereof, retrieves personal information. In at least some embodiments, the controller analyzes the feature vectors to detect personal information about the occupant related to the live prompt. In at least some embodiments, the controller determines which personal information is relevant by comparing the feature vectors calculated from the live prompt with the feature vectors of the personal information. In at least some embodiments, the controller determines relevant personal information by the proximity of the feature vectors of the personal information to the feature vectors calculated from the raw prompt in feature space. In at least some embodiments, the controller applies a distance formula to calculate the distance between the feature vectors in feature space and identifies the personal information corresponding to the feature vector closest to the feature vector calculated from the raw prompt. In at least some embodiments, the controller uses Retrieval Augmented Generation (RAG) techniques.

[0056] At S646, the controller, or a portion thereof, adds content associated with the live prompt. In at least some embodiments, the controller adds content associated with the live prompt based on the identified personal information. In at least some embodiments, this action results in the addition of personal information associated with the live prompt. In at least some embodiments, the controller performs this process to increase the efficiency and accuracy of the output of the LLM in response to the prompt.

[0057] At S648, the controller, or a portion thereof, frames the added prompt. In at least some embodiments, the controller structures the added prompt into a format suitable for input to the LLM. In at least some embodiments, the added and framed prompt is used as a pre-edit prompt in the cloud-assisted in-vehicle large-scale language model. In at least some embodiments, the controller uses a processing prompt frame. In at least some embodiments, the controller searches the LLM for context through the processing prompt frame. In at least some embodiments, the controller uses a prompt frame to inform the LLM that the user is driving the vehicle.

[0058] 7 is a schematic diagram of a vehicle according to at least some embodiments of the present disclosure. The system includes a vehicle 700, a first LLM 710, personal information 712, a raw prompt 750, a search query generator 718, a query prompt 752, a search query 754, a feature vector calculator 719, a feature vector 756, related content 758, and a pre-edit prompt 720. In at least some embodiments, the vehicle 700 is configured to use RAG technology.

[0059] The vehicle 700, first LLM 710, personal information 712, and pre-edit prompt 720 are substantially similar in structure and function to the vehicle 100, first LLM 110, personal information 112, and pre-edit prompt 120, respectively, unless otherwise specified.

[0060] The raw prompt 750 is the first detected input from the user. In at least some embodiments, the raw prompt 750 is a string of characters, a voice command, etc. In at least some embodiments, the raw prompt 750 is detected by a microphone, an image sensor, etc.

[0061] The search query generator 718 is configured to generate a query prompt based on the raw prompt. In at least some embodiments, the search query generator 718 is configured to generate a query prompt 752 based on the raw prompt 750. In at least some embodiments, the search query generator 718 is configured to input the query prompt 752 to the first LLM 710 and cause the first LLM 710 to output a search query 754. In at least some embodiments, the search query generator 718 is a software module, an algorithm, etc. In at least some embodiments, the search query generator 718 is configured to receive the search query 754 from the first LLM 710. In at least some embodiments, the search query generator 718 is configured to pass the search query 754 to the feature vector calculator 719.

[0062] The query prompt 752 is generated by a search query generator. In at least some embodiments, the query prompt 752 is configured as a prompt to cause the first LLM 710 to output a search query to identify relevant personal information. In at least some embodiments, the query prompt 752 is a string representing the search query.

[0063] The search query 754 is output by the first LLM 710 to assist in identifying personal information. In at least some embodiments, the search query 754 is configured to be combined with the live prompt 750.

[0064] Feature vector calculator 719 is configured to calculate a feature vector based on the search query. In at least some embodiments, feature vector calculator 719 is configured to calculate a feature vector based on the search query 754 and the live prompt 750. In at least some embodiments, feature vector calculator 719 is configured to apply a machine learning model to calculate feature vector 756.

[0065] Feature vector 756 is used to identify relevant content within personal information 712. In at least some embodiments, feature vector 756 is a numeric vector. In at least some embodiments, feature vector 756 includes embedded information for associating with personal information. In at least some embodiments, feature vector 756 represents the location of search query 754 in feature space.

[0066] Related content 758 represents personal information related to live prompt 750. In at least some embodiments, related content 758 is combined with live prompt 750 to form pre-edit prompt 720. In at least some embodiments, related content 758 is text, multimedia content, etc.

[0067] 8 is an operational flow diagram for logit adjustment in accordance with at least some embodiments of the present disclosure. In at least some embodiments, the operational flow diagram presents a method for logit adjustment in accordance with at least some embodiments of the present disclosure. In at least some embodiments, the method is performed by a controller of a server, such as controller 1002 of server 1000 in FIG. 10, described below.

[0068] At S860, the controller, or a portion thereof, redacts personal information from the pre-edit prompt. In at least some embodiments, the controller removes personal information from the pre-edit prompt to enhance privacy. In at least some embodiments, the controller executes the operational flow of FIG. 9, described below.

[0069] At S862, the controller, or a portion thereof, passes the edited prompt to a second large-scale language model (LLM). In at least some embodiments, the controller passes the edited prompt to a second LLM located on a cloud network. In at least some embodiments, the controller causes a transmitter on the vehicle to transmit the edited prompt over a network, such as a Wi-Fi network or a wireless access network.

[0070] At S864, the controller, or a portion thereof, receives the logits output from the second LLM. In at least some embodiments, the controller receives the logits output from the second LLM from the cloud network. In at least some embodiments, the controller collects the logits output from the second LLM. In at least some embodiments, the controller instructs the cloud network to transmit the logits output from the second LLM.

[0071] At S865, the controller, or a portion thereof, inputs the post-edit prompt to a third LLM. In at least some embodiments, the controller inputs the post-edit prompt to a third LLM operating in the vehicle. In at least some embodiments, the controller causes the post-edit prompt to be input to a third LLM operating in the cloud network.

[0072] At S867, the controller, or a portion thereof, calculates a difference between the logits of the first LLM and the third LLM. In at least some embodiments, the controller calculates the difference between corresponding beliefs in the logits of the first LLM and the third LLM. In at least some embodiments, the controller calculates the difference by comparing the logits of the first LLM and the third LLM. In at least some embodiments, the controller generates a difference having the same dimensions as the logits of the first LLM and the third LLM.

[0073] At S869, the controller, or a portion thereof, calculates a sum of the difference and the logit of the second LLM. In at least some embodiments, the controller calculates the sum of the difference and the logit of the second LLM. In at least some embodiments, the controller calculates the sum to obtain an adjusted logit. In at least some embodiments, the controller generates a sum having the same dimensions as the logits of the first, second, and third LLMs.

[0074] 9 is an operational flow diagram for prompt editing, in accordance with at least some embodiments of the present disclosure. In at least some embodiments, the operational flow diagram presents a method of prompt editing, in accordance with at least some embodiments of the present disclosure. In at least some embodiments, the method is performed by a controller of a server, such as controller 1002 of server 1000 of FIG. 10 , described below.

[0075] At S970, the controller, or a portion thereof, detects personal information. In at least some embodiments, the controller scans the pre-edit prompt to identify personal information. In at least some embodiments, the controller compares the pre-edit prompt to a personal information database to identify personal information. In at least some embodiments, the controller performs pattern recognition, keyword detection, or other data analysis techniques. In at least some embodiments, the controller identifies one or more portions of the pre-edit prompt that contain personal information.

[0076] At S972, the controller, or a portion thereof, selects a method for deleting the personal information. In at least some embodiments, the controller selects an appropriate deletion method based on the type and context of the detected personal information. In at least some embodiments, the controller selects a method from anonymization, pseudonymization, or full deletion. In at least some embodiments, the controller selects a method that balances privacy and accuracy. In at least some embodiments, the controller selects a method that maintains the integrity of the prompt while protecting the user's personal information.

[0077] At S974, the controller, or a portion thereof, applies the selected personal information deletion method. In at least some embodiments, the controller applies the selected deletion method to the pre-edit prompt. In at least some embodiments, this action effectively redacts the identified personal information. In at least some embodiments, this action does not result in any action other than the prompt edit action. In at least some embodiments, this action is performed after the deletion method is selected. In at least some embodiments, the result of this action is a post-edit prompt with the personal information removed. In at least some embodiments, this action creates a version of the prompt that respects the user's privacy.

[0078] At S976, the controller, or a portion thereof, determines whether all personal information has been deleted. In response to a determination that all personal information has not been completely deleted, the operational flow returns to detecting personal information at S970. In response to a determination that all personal information has been deleted, the operational flow ends. In at least some embodiments, the controller validates the post-edit prompt to ensure that all personal information has been properly deleted.

[0079] FIG. 10 is a schematic diagram of a hardware configuration for use of cloud-assisted in-vehicle large-scale language models, in accordance with at least some embodiments of the present disclosure.

[0080] A typical hardware configuration includes a server 1000 that interacts with an input device 1007, either directly or through a cloud network 1009. In at least some embodiments, the input device 1007 is a touchscreen, microphone, camera, or other device configured to detect tactile, auditory, visual, or other input. In at least some embodiments, the cloud network 1009 is a wide area network, such as the Internet. In at least some embodiments, the server 1000 is a computer or other computing device that receives input or commands from the input device 1007. In at least some embodiments, the server 1000 is integrated with the input device 1007. In at least some embodiments, the server 1000 is a computer system that executes computer-readable instructions to perform operations for use with the cloud-assisted in-vehicle large-scale language model.

[0081] The server 1000 includes a controller 1002, storage 1004, an input / output interface 1006, and a communication interface 1008. In at least some embodiments, the controller 1002 includes a processor or programmable circuit that executes instructions to perform operations according to the instructions. In at least some embodiments, the controller 1002 includes analog or digital programmable circuitry, or a combination thereof. In at least some embodiments, the controller 1002 includes physically separate storage or circuitry that interacts through communications. In at least some embodiments, the storage 1004 includes a non-transitory computer-readable medium capable of storing executable and non-executable data accessed by the controller 1002 during execution of instructions. In at least some embodiments, the communication interface 1008 transmits and receives data to and from a cloud network 1009. In at least some embodiments, input / output interface 1006 connects various input and output units, such as input devices 1007, via parallel ports, serial ports, keyboard ports, mouse ports, monitor ports, etc., to accept commands and present information. In some embodiments, storage 1004 is external to server 1000.

[0082] The controller 1002 includes an input unit 1080, an adjustment unit 1082, an editing unit 1084, and a calculation unit 1086. The storage 1004 includes personal information 1090, a large-scale language model 1092, adjustment parameters 1094, and editing parameters 1096.

[0083] Input 1080 is circuitry or instructions in controller 1002 configured to input prompts to the LLMs. In at least some embodiments, input 1080 is configured to input pre-edit prompts to a first LLM implemented in the vehicle, input post-edit prompts to a second LLM implemented in the cloud network, and input post-edit prompts to a third LLM implemented in the vehicle or the cloud network. In at least some embodiments, input 1080 uses information in storage 1004, such as large-scale language models 1092. In at least some embodiments, input 1080 includes sub-portions for performing additional functions, such as those described in the preceding flowcharts. In at least some embodiments, such sub-portions are referred to by names related to their corresponding functions.

[0084] Adjuster 1082 is circuitry or instructions in controller 1002 configured to adjust the logits. In at least some embodiments, adjuster 1082 calculates the adjusted logits based on the logits output from the first LLM, the second LLM, and the third LLM. In at least some embodiments, adjuster 1082 uses information in storage 1004, such as adjustment parameters 1094. In at least some embodiments, adjuster 1082 includes subparts for performing additional functions, such as those described in the preceding flowcharts. In at least some embodiments, such subparts are referred to by names related to their corresponding functions.

[0085] The editor 1084 is circuitry or instructions in the controller 1002 configured to edit the pre-edit prompt. In at least some embodiments, the editor 1084 is configured to edit the pre-edit prompt to remove personal information to create a post-edit prompt. In at least some embodiments, the editor 1084 uses information in the storage 1004, such as personal information 1090 and editing parameters 1096. In at least some embodiments, the editor 1084 includes sub-portions for performing additional functions, such as those described in the preceding flowcharts. In at least some embodiments, such sub-portions are referred to by names related to their corresponding functions.

[0086] The calculator 1086 is circuitry or instructions in the controller 1002 configured to calculate the difference between the logits. In at least some embodiments, the calculator 1086 is configured to calculate the difference between the logits of the first LLM and the logits of the third LLM. In at least some embodiments, the calculator 1086 uses information in the storage 1004, such as the tuning parameters 1094. In at least some embodiments, the calculator 1086 includes subdivisions for performing additional functions, such as those described in the preceding flowcharts. In at least some embodiments, such subdivisions are referred to by names related to their corresponding functions.

[0087] In at least some embodiments, the apparatus is another device capable of processing logical functions to perform the operations described herein. In at least some embodiments, the controller and storage unit may not be completely separate devices, but in some embodiments share circuitry or one or more computer-readable media. In at least some embodiments, the storage unit stores both computer-executable instructions and data accessed by the controller, which includes a central processing unit (CPU) and RAM combination, where all or a portion of the computer-executable instructions are replicable to the CPU for execution during the performance of the operations described herein.

[0088] In at least some embodiments, the apparatus is a computer, and programs installed on the computer can cause the computer to function as the apparatus of the embodiments described herein or to perform operations associated with the apparatus of the embodiments described herein. In at least some embodiments, such programs can be executed by a processor to cause the computer to perform predetermined operations associated with some or all of the blocks in the flowcharts and block diagrams described herein.

[0089] At least some embodiments are described with reference to flowcharts and block diagrams with blocks representing (1) method steps by which operations are performed or (2) portions of a controller responsible for performing operations. In at least some embodiments, certain steps and portions are implemented by dedicated circuitry, programmable circuitry with computer-readable instructions stored on a computer-readable medium, and / or a processor with computer-readable instructions stored on a computer-readable medium. In at least some embodiments, dedicated circuitry includes digital and / or analog hardware circuitry, integrated circuits (ICs), and / or discrete circuitry. In at least some embodiments, programmable circuitry includes reconfigurable hardware circuitry including logical AND, OR, XOR, NAND, NOR, and other logic operations, flip-flops, registers, memory elements, etc., such as field-programmable gate arrays (FPGAs), programmable logic arrays (PLAs), etc.

[0090] In at least some embodiments, a computer-readable storage medium is a tangible device capable of storing and maintaining instructions for use by an instruction execution device. In some embodiments, a computer-readable storage medium includes, for example, but not limited to, electronic, magnetic, optical, electromagnetic, or semiconductor storage devices, or suitable combinations thereof. A non-exhaustive list of more specific examples of computer-readable storage media includes: portable computer diskettes, hard disks, Random Access Memory (RAM), Read-Only Memory (ROM), Erasable Programmable Read-Only Memory (EPROM; flash memory), Static Random Access Memory (SRAM), Compact Disc Read-Only Memory (CD-ROM), Digital Versatile Disk (DVD), Memory Stick, floppy disks, mechanically encoded devices such as punch cards or grooved structures having instructions stored thereon, and suitable combinations thereof. The computer-readable medium referred to herein should not be construed as a transitory signal itself, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through a waveguide or other transmission medium (e.g., light pulses through a fiber optic cable), or electrical signals transmitted over electrical wires.

[0091] In at least some embodiments, the computer-readable program instructions described herein can be downloaded from a computer-readable storage medium to a separate computing / processing device or to an external computer or external storage device over a network, e.g., the Internet, a local area network, a wide area network, and / or a wireless network. In at least some embodiments, the network includes copper transmission cables, optical transmission cables, wireless transmissions, routers, firewalls, switches, gateway computers, and / or edge servers. In at least some embodiments, a network adapter card or network interface on each computing / processing device receives the computer-readable program instructions from the network and forwards the computer-readable program instructions for storage on a computer-readable storage medium within the separate computing / processing device.

[0092] In at least some embodiments, the computer-readable program instructions for performing the operations described above are assembler instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, state-setting data, or other source or object code written in one or more of a combination of programming languages, including object-oriented programming languages ​​such as Smalltalk, C++, etc., and traditional procedural programming languages ​​such as the "C" programming language or similar programming languages. In at least some embodiments, the computer-readable program instructions execute entirely on the user's computer as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In at least some embodiments, in the latter scenario, the remote computer is connected to the user's computer through some type of network, including a local area network (LAN), a wide area network (WAN), or a connection to an external computer is made (e.g., over the Internet using an Internet connection provider). In at least some embodiments, electronic circuitry including, for example, a programmable circuit, a field-programmable gate array (FPGA), or a programmable logic array (PLA) uses the state information of the computer-readable program instructions to execute the computer-readable program instructions to individualize the electronic circuitry and perform aspects of the present invention.

[0093] Although the embodiments of the present invention have been described above, the technical scope of the claimed subject matter is not limited to the above-described embodiments. Those skilled in the art will understand that various modifications and improvements to the above-described embodiments are possible. Furthermore, those skilled in the art will understand from the claims that embodiments to which such modifications or improvements have been made are included within the technical scope of the invention.

[0094] The operations, procedures, steps, and stages of each process performed in the devices, systems, programs, and methods shown in the claims, embodiments, or drawings may be performed in any order unless the order is indicated as "before" or "after," and the output of a previous process is not used in a subsequent process. Even if a process flow is described in the claims, embodiments, or drawings using words such as "first" or "next," such a description does not necessarily mean that the process must be performed in the order described.

[0095] In at least some embodiments, the use of cloud-assisted in-vehicle large-scale language models is performed by inputting pre-edit prompts to a first large-scale language model (LLM) implemented in the vehicle, the pre-edit prompts including personal information related to an occupant of the vehicle, sending post-edit prompts to a second LLM implemented in the cloud, the post-edit prompts representing the pre-edit prompts without the personal information, receiving logits of the second LLM, inputting the post-edit prompts to a third LLM implemented in the cloud or in the vehicle, the third LLM including the same parameter set as the first LLM, calculating a difference between the logits of the first LLM and the logits of the third LLM, calculating a sum of the difference and the logits of the second LLM to obtain adjusted logits, and obtaining an output from the adjusted logits.

[0096] In at least some embodiments, the number of parameters of the second LLM is greater than the number of parameters of each of the first LLM and the third LLM. In at least some embodiments, the first LLM is trained using personal information, and the second LLM and the third LLM are trained without personal information. In at least some embodiments, using the cloud-assisted in-vehicle large-scale language model includes providing an output to an occupant or using the output to control the vehicle. In at least some embodiments, using the cloud-assisted in-vehicle large-scale language model further includes detecting a live prompt, the live prompt including at least one of a voice sample or an image of the occupant. In at least some embodiments, using the cloud-assisted in-vehicle large-scale language model further includes adding predetermined information to the live prompt. In at least some embodiments, the predetermined information includes processed prompt information. In at least some embodiments, using the cloud-assisted in-vehicle large-scale language model further includes calculating additional occupant information associated with the live prompt. In at least some embodiments, the pre-edited prompt includes the additional occupant information. In at least some embodiments, the third LLM includes the same parameter values ​​as the first LLM. In at least some embodiments, the third LLM is the first LLM. In at least some embodiments, obtaining an output includes summing multiple iteration outputs, each iteration output corresponding to repeatedly obtaining adjusted logits in a multiple iteration sequence, and the pre-editing prompt in each iteration includes each iteration output corresponding to a previous iteration in the multiple iteration sequence. In at least some embodiments, using the cloud-assisted in-vehicle large-scale language model further includes editing personal information from the pre-editing prompt to create a post-editing prompt. In at least some embodiments, editing includes at least one of converting speech to text, converting image data to object data, generalizing location, filtering the raw prompt, or randomizing the raw prompt. In at least some embodiments, each logit has the same dimensionality.

[0097] In at least some embodiments, the use of the cloud-assisted in-vehicle large-scale language model is performed by a device that includes a processor that executes instructions according to the operations described above, or a controller that includes circuitry configured to perform the operations described above.

[0098] With the foregoing overview of features of some embodiments, those skilled in the art will have a better understanding of the aspects of the present disclosure. Those skilled in the art will appreciate that the present disclosure may readily serve as a basis for the design or modification of other processes and structures to carry out the same purposes and / or achieve similar effects as the embodiments presented herein. Those skilled in the art will also realize that such equivalents do not depart from the spirit and scope of the present disclosure, and that various changes, substitutions, and adjustments can be made to the present disclosure without departing from the spirit and scope of the present disclosure.

Claims

1. inputting pre-edit prompts into a first large scale language model (LLM) implemented in a vehicle, the pre-edit prompts including personal information associated with an occupant of the vehicle; sending a post-edit prompt to a second LLM implemented in the cloud, the post-edit prompt representing the pre-edit prompt without the personal information; receiving the logits of the second LLM; inputting the post-edit prompt into a third LLM implemented in the vehicle or the cloud, the third LLM including the same parameter set as the first LLM; calculating the difference between the logit of the first LLM and the logit of the third LLM; calculating a sum of the difference and the logit of the second LLM to obtain an adjusted logit; obtaining an output from the adjusted logit; A computer program that causes one or more processors to perform operations including:

2. The computer program product of claim 1 , wherein the number of parameters of the second LLM is greater than the number of parameters of each of the first LLM and the third LLM.

3. the first LLM is trained using the personal information; The second LLM and the third LLM are trained without using the personal information.

3. A computer program according to claim 1 or 2.

4. The computer program product of claim 1 or 2, wherein the action further comprises providing the output to the occupant or using the output to control the vehicle.

5. The computer program product of claim 1 or 2, wherein the actions further comprise detecting a live prompt, the live prompt comprising at least one of a voice sample or an image of the occupant.

6. The computer program product of claim 5 , wherein the action further comprises adding predetermined information to the live prompt.

7. The computer program product of claim 6 , wherein the predetermined information includes processing prompt information.

8. The computer program product of claim 5 , wherein the actions further include calculating additional occupant information associated with the live prompt.

9. The computer program product of claim 8 , wherein the pre-edit prompt includes the additional passenger information.

10. The computer program of claim 1 or 2, wherein the third LLM includes the same parameter values ​​as the first LLM.

11. The computer program of claim 1 or 2, wherein the third LLM is the first LLM.

12. obtaining the output includes summing the multiple iteration outputs; Each iteration output corresponds to repeatedly obtaining adjusted logits over multiple iteration sequences; the pre-edit prompt in each iteration includes a respective iteration output corresponding to a preceding iteration in the sequence of the plurality of iterations; 3. A computer program according to claim 1 or 2.

13. The computer program product of claim 5 , wherein the actions further include editing the personal information from the pre-edit prompt to create the post-edit prompt.

14. 14. The computer program product of claim 13, wherein the editing comprises at least one of converting speech to text, converting image data to object data, generalizing a position, filtering the live prompt, or randomizing the live prompt.

15. 3. The computer program of claim 1, wherein each logit has the same dimension.

16. 1. A method executed by one or more processors, comprising: inputting pre-edit prompts into a first large scale language model (LLM) implemented in a vehicle, the pre-edit prompts including personal information associated with an occupant of the vehicle; sending a post-edit prompt to a second LLM implemented in the cloud, the post-edit prompt representing the pre-edit prompt without the personal information; receiving the logits of the second LLM; inputting the post-edit prompt into a third LLM implemented in the vehicle or the cloud, the third LLM including the same parameter set as the first LLM; calculating the difference between the logit of the first LLM and the logit of the third LLM; calculating a sum of the difference and the logit of the second LLM to obtain an adjusted logit; obtaining an output from the adjusted logit; The method includes:

17. 17. The method of claim 16, wherein the number of parameters of the second LLM is greater than the number of parameters of each of the first LLM and the third LLM.

18. the first LLM is trained using the personal information; The second LLM and the third LLM are trained without using the personal information.

18. The method of claim 16 or 17.

19. 18. The method of claim 16 or 17, further comprising providing said output to said occupant or using said output to control said vehicle.

20. inputting pre-edit prompts into a first large scale language model (LLM) implemented in a vehicle, the pre-edit prompts including personal information associated with an occupant of the vehicle; sending a post-edit prompt to a second LLM implemented in the cloud, the post-edit prompt representing the pre-edit prompt without the personal information; receiving the logits of the second LLM; inputting the post-edit prompt into a third LLM implemented in the vehicle or the cloud, the third LLM including the same parameter set as the first LLM; calculating the difference between the logit of the first LLM and the logit of the third LLM; calculating a sum of the difference and the logit of the second LLM to obtain an adjusted logit; obtaining an output from the adjusted logit; 10. An apparatus comprising: a controller including circuitry configured to perform operations including:

Images