Cloud-assisted large-scale in-vehicle 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 Art

[0001] To improve the user experience in vehicles, a Large Language Model (LLM) is used to perform vehicle control, provide information suitable for passengers, etc. The LLM can be executed using local computing resources or cloud-based computing resources.

Brief Description of the Drawings

[0002] Each aspect of the present disclosure is best understood by reading the following embodiments for carrying out the invention together with the accompanying drawings. Note that the various drawings are not drawn to scale according to the convention in the industry. In fact, the dimensions in the various drawings can be arbitrarily enlarged or reduced for clarity of explanation.

[0003] [Figure 1] FIG. 1 is a schematic diagram of a system for a cloud-assisted in-vehicle large language model in at least one embodiment of the present disclosure.

[0004] [Figure 2] FIG. 2 is a schematic diagram of a system for a cloud-assisted in-vehicle large language model in at least one embodiment of the present disclosure.

[0005] [Figure 3] FIG. 3 is a schematic diagram of a system for a cloud-assisted in-vehicle large language model in at least one embodiment of the present disclosure.

[0006] [Figure 4] FIG. 4 is a schematic diagram of a system for a cloud-assisted in-vehicle large language model in at least one embodiment of the present disclosure.

[0007] [Figure 5] FIG. 5 is an operation flowchart of a system for a cloud-assisted in-vehicle large language model in at least one embodiment of the present disclosure.

[0008] [Figure 6] Figure 6 is a flowchart illustrating the operation for preparing a pre-edit prompt in at least one embodiment of the present disclosure.

[0009] [Figure 7] Figure 7 is a schematic diagram of a vehicle in at least one embodiment of the present disclosure.

[0010] [Figure 8] Figure 8 is a flowchart of the operation for logit adjustment in at least one embodiment of the present disclosure.

[0011] [Figure 9] Figure 9 is an operational flowchart for prompt editing in at least one embodiment of the present disclosure.

[0012] [Figure 10] Figure 10 is a schematic diagram of a hardware configuration for using a cloud-assisted in-vehicle large-scale language model in at least one embodiment of the present disclosure. [Modes for carrying out the invention]

[0013] The following disclosure provides numerous different embodiments or examples to implement different features of the subject matter presented. For the sake of simplicity, specific examples of components, values, operations, materials, combinations, etc. are described below. These are, of course, merely examples and not intended to be limiting. Other components, values, operations, materials, combinations, etc. are also considered. In addition, reference figures and / or reference numerals may be repeated in different examples in this disclosure. This repetition is for simplicity and clarity and does not in itself indicate a relationship between the different embodiments and / or configurations in which it is done.

[0014] LLMs implemented in vehicles may have limitations in output accuracy due to the constraints of the vehicle's capacity and processing. Using LLMs implemented in the cloud may increase the risk of leakage of personal information, including personal information of occupants, included in the input to the LLM. Also, since output from inputs containing personal information may be more accurate than output from inputs without personal information, removing such personal information from the input may limit the accuracy of the output.

[0015] In at least one embodiment of this disclosure, a more accurate output can be obtained 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 a larger LLM to obtain an adjusted logit. In at least one embodiment, the logit is an intermediate output of the LLM, represented in the form of a denormalized vector containing values ​​representing the confidence level of the most appropriate candidate for output. For example, if the prompt input to the LLM is "What is the material of this road?", the logit may include values ​​representing a confidence level of 95% for "asphalt", 94.6% for "concrete", 91.2% for "soil", and 84.3% for "gravel". In at least one embodiment, the difference in logits represents the mathematical difference of the values ​​representing the confidence levels of each common candidate. For example, if the smaller LLM logit from the pre-edit prompt includes a value representing a 95% confidence level for "asphalt," and the smaller LLM logit from the post-edit prompt includes a value representing a 94.7% confidence level for "asphalt," the difference in values ​​represents an increase of approximately 0.3% in the confidence level for "asphalt." Note that the conversion from denormalized logit values ​​to confidence percentages does not yield such an exact difference in confidence. In at least one embodiment, this increase in confidence is presumed to be due to the inclusion of personal information in the input. In at least one embodiment, applying the difference to the logit represents the mathematical sum of the difference value and the logit value representing the confidence level of the candidate. For example, if the larger LLM logit from the post-edit prompt includes a value representing a 94.2% confidence level for "asphalt," applying the difference value representing a 0.3% increase results in an adjusted logit representing a 94.5% confidence level for "asphalt." It should be noted here that the conversion from unnormalized logit values ​​to confidence percentages does not yield an exact sum of such confidence levels. In at least one embodiment, the best candidate according to the adjusted logit may differ from the best candidate according to the logit of the smaller LLM from the pre-edit prompt and the best candidate according to the logit of the larger LLM from the post-edit prompt.

[0016] In at least one embodiment, using a large LLM in the cloud without sending personal information to the cloud results in more accurate output from the vehicle's LLM. In at least one embodiment, using other smaller LLMs in the cloud reduces the computational load on the vehicle.

[0017] In at least one embodiment, all LLMs have logits of the same dimension. In at least one embodiment, one or two smaller LLMs have the same set of parameters but different parameter values. In at least one embodiment, one of the two smaller LLMs is learned for personal information, and the other of the two smaller LLMs, as well as the larger LLM, are learned for general information. In at least one embodiment, the difference between logits is calculated from the logit from the pre-edit prompt and the logit from the post-edit prompt output from one of the smaller LLMs. In at least one embodiment, resource consumption can be reduced by using one of the smaller LLMs to calculate the difference.

[0018] In at least one embodiment, the vehicle has at least one small LLM, and the large LLM is in the cloud. In at least one embodiment, other small LLMs may be contained in either the vehicle or the cloud. In at least one embodiment, 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 either the vehicle or the cloud. In at least one embodiment, if other commonly learned small LLMs are in the cloud, it makes sense to perform the comparison in the cloud. In at least one embodiment, the difference and the logit of the large LLM are added in either the vehicle or the cloud. In at least one embodiment, the iterative outputs are summed in either the vehicle or the cloud.

[0019] In at least one embodiment, the raw prompt may be an instruction, inquiry, command, etc. In at least one embodiment, the raw prompt may be embedded in a processed prompt that provides context for creating the pre-edited prompt. For example, the processed prompt may add predetermined information, such as information that the user is driving a vehicle. In at least one embodiment, the raw prompt may be supplemented with personal information relevant to the raw prompt, such as precise location, allergy information, recent behavior, known preferences, etc. In at least one embodiment, Retrieval Augmented Generation (RAG) technology may be used to determine the relevant personal information. Examples of personal information, such as occupant information, include facial images of occupants (e.g., drivers), voice recordings of occupants, identification information of occupants' mobile devices (e.g., MAC addresses), vehicle operations by occupants, and information about occupants pre-stored in a storage device. In at least one embodiment, the pre-edited prompt is edited into a post-edited prompt by means of text-to-speech conversion, image data-to-object data conversion, location generalization, filtering of raw prompts, randomization of raw prompts, etc.

[0020] Figure 1 is a schematic diagram of a system for a cloud-assisted in-vehicle large-scale language model in at least one embodiment of the present disclosure. The system includes a vehicle 100, a cloud network 109, a first LLM 110, a pre-edit prompt 120, personal information 112, a first LLM logit 121, a second LLM 114, a post-edit prompt 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 a first LLM 110 and personal information 112, and communicates with a cloud network 109. In at least some embodiments, Vehicle 100 is configured to support the installation and operation of the first LLM 110. In at least some embodiments, Vehicle 100 is configured to provide transportation to its passengers. In at least some embodiments, Vehicle 100 is configured to input a pre-editing prompt 120 to the first LLM 110. In at least some embodiments, Vehicle 100 is configured to receive a third LLM logit 126 from the cloud network 109. In at least some embodiments, Vehicle 100 is configured to calculate a logit difference 127 by comparing the first LLM logit 121 and the third LLM logit 126. In at least some embodiments, Vehicle 100 is configured to receive a second LLM logit 124 from the cloud network 109. In at least some embodiments, Vehicle 100 is configured to calculate a logit sum 128 by adding the logit difference 127 to the second LLM logit 124. In at least some embodiments, Vehicle 100 is configured to transmit the personal information 112 to the first LLM 110, including a network such as a CAN (Controller Area Network). In at least some embodiments, Vehicle 100 is one of a number of vehicles equipped with an in-vehicle computing system. In at least some embodiments, Vehicle 100 is an automobile, a truck, a boat, an airplane, a submarine, etc. In at least some embodiments, Vehicle 100 is configured for transportation, navigation, entertainment, etc.

[0022] The cloud network 109 communicates with the vehicle 100 and includes a second LLM 114 and a third LLM 116. In at least one embodiment, the cloud network 109 is configured to transmit data between the vehicle 100 and the second LLM 114 and the third LLM 116. In at least one embodiment, the cloud network 109 is configured to receive an edited prompt 123 from the vehicle 100 and to input the edited prompt 123 to the second LLM 114 and the third LLM 116. In at least one embodiment, the cloud network 109 is configured to transmit a second LLM logit 124 and a third LLM logit 126 to the vehicle 100. In at least one embodiment, the cloud network 109 is configured to provide a platform for various cloud-based services. In at least one embodiment, the cloud network 109 is configured to interact with an internet service provider and a data center. In at least one embodiment, the cloud network 109 includes a data center having servers that host cloud services. In at least one embodiment, the cloud network 109 is configured for data storage, web hosting, SaaS (Software as a Service), and other purposes.

[0023] The first LLM 110 is included in the vehicle 100 and is configured to process a prompt and generate logits. In at least some embodiments, the first LLM 110 is configured to process the pre-editing prompt 120 and generate the first LLM logits 121. In at least some embodiments, the first LLM 110 is configured to interact with the personal information 112. In at least some embodiments, the first LLM 110 is 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 is trained using at least a portion of the 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 in the computing system of the vehicle 100. In at least some embodiments, the first LLM 110 includes a neural network, instructions for processing inputs through the neural network, and instructions for converting logits into an output interpretable by a human or a computer.

[0024] The pre-editing prompt 120 is the first input to the first LLM 110 and serves as the basis for the post-editing prompt 123. In at least some embodiments, the pre-editing prompt 120 includes personal information such as information within the personal information 112. In at least some embodiments, the pre-editing prompt 120 is used as a basis for creating the post-editing prompt 123. In at least some embodiments, the pre-editing prompt 120 is a voice command, text input, etc. In at least some embodiments, the pre-editing prompt 120 includes inquiries, commands, etc. from the user.

[0025] Personal information 112 is collected personal information, such as personal information contained in the pre-edit prompt 120. In at least one embodiment, personal information 112 is used to edit the personal information contained in the pre-edit prompt 120 to create the post-edit prompt 123. In at least one embodiment, personal information 112 is specific to the occupants of one or more vehicles. In at least one embodiment, personal information 112 is used to append to the raw prompt. In at least one embodiment, personal information 112 includes a database that stores and indexes the collected personal information. In at least one embodiment, the collected personal information includes the user's name, preferences, etc.

[0026] The first LLM logit 121 is the output of the first LLM 110 based on the pre-edit prompt 120 and is used to calculate the logit difference 127. In at least one embodiment, the first LLM logit 121 represents the intermediate output of the first LLM 110. In at least one embodiment, the first LLM logit 121 is a vector of probabilities for each word that could be the next word in the word sequence. In at least one embodiment, the first LLM logit 121 represents the confidence level of each of the most likely candidates as the output.

[0027] The second LLM 114 resides within the cloud network 109 and is configured to process prompts and create a logit. In at least one embodiment, the second LLM 114 is configured to process a post-edit prompt 123 and create a second LLM logit 124. In at least one embodiment, the second LLM 114 has a larger parameter set than the first LLM 110. In at least one embodiment, the second LLM 114 has more layers than the first LLM 110. In at least one embodiment, the second LLM 114 has more interconnections than the first LLM 110. In at least one embodiment, the second LLM 114 is trained for various language modeling tasks such as natural language processing, speech recognition, and image processing. In at least one embodiment, the second LLM 114 is trained for general use without using personal information 112. In at least one embodiment, the second LLM 114 is configured to interact with vehicles and computers different from vehicle 100. In at least one embodiment, the second LLM 114 includes a software module that runs using the computing resources of the cloud network 109. In at least one embodiment, the second LLM 114 includes a neural network, instructions for processing input through the neural network, and instructions for translating logits into human or computer-interpretable outputs.

[0028] The edited prompt 123 represents the pre-edited prompt 120 from which personal information has been removed. In at least one embodiment, the edited prompt 123 is input to the second LLM 114 and the third LLM 116. In at least one embodiment, the edited prompt 123 undergoes processing from the pre-edited prompt 120, such as text-to-speech conversion, image data-to-object data conversion, location generalization, filtering, and randomization.

[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 one embodiment, the second LLM logit 124 represents the intermediate output of the second LLM 114. In at least one embodiment, the second LLM logit 124 is a vector of probabilities for each word that could be the next word in the word sequence. In at least one embodiment, the second LLM logit 124 represents the confidence level of each of the most likely candidates as the output. In at least one embodiment, the dimension of the second LLM logit 124 is identical to that of the first LLM logit 121.

[0030] The third LLM 116 resides within the cloud network 109 and is configured to process prompts and create a logit. In at least one embodiment, the third LLM 116 is configured to process a post-edit prompt 123 and create a third LLM logit 126. In at least one embodiment, the third LLM 116 has the same parameter set as the first LLM 110. In at least one embodiment, the third LLM 116 has different parameter values ​​than the first LLM 110. In at least one embodiment, the third LLM 116 is trained for various language modeling tasks such as natural language processing, speech recognition, and image processing. In at least one embodiment, the third LLM 116 is trained for general use without using personal information 112. In at least one embodiment, the third LLM 116 is configured to interact with vehicles and computers different from vehicle 100. In at least one embodiment, the third LLM 116 includes a software module that runs using the computing resources of the cloud network 109. In at least one embodiment, the third LLM116 includes a neural network, instructions for processing input through the neural network, and instructions for translating logits into human or computer-interpretable outputs.

[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 one embodiment, the third LLM logit 126 represents the intermediate output of the third LLM 116. In at least one embodiment, the third LLM logit 126 is a vector of probabilities for each word that could be the next word in the word sequence. In at least one embodiment, the third LLM logit 126 represents the confidence level of each of the most likely candidates as the output. In at least one embodiment, the dimensions of the third LLM logit 126 are the same as those of the first LLM logit 121 and the second LLM logit 124.

[0032] The logit difference 127 represents the difference between the first LLM logit 121 and the third LLM logit 126. In at least one embodiment, the logit difference 127 is used to calculate the logit sum 128. In at least one embodiment, the logit difference 127 is a vector of differences in the probability of each word being the next word in the word sequence. In at least one embodiment, the logit difference 127 represents the difference in the confidence level of each of the most likely candidates as the output. In at least one embodiment, the dimension of the logit difference 127 is the same as that of the first LLM logit 121, the second LLM logit 124, and the third LLM logit 126.

[0033] The logit sum 128 represents the sum of the logit difference 127 and the second LLM logit 124. In at least one embodiment, the logit sum 128 represents the adjustment logit. In at least one embodiment, the logit sum 128 is the adjustment logit used to obtain the output. In at least one embodiment, the logit sum 128 is a vector of probabilities for each word that could be the next word in the word sequence. In at least one embodiment, the logit sum 128 represents the confidence level of each of the most likely candidates as the output. In at least one embodiment, the dimension of the logit sum 128 is the same as that of the first LLM logit 121, the second LLM logit 124, and the third LLM logit 126.

[0034] Figure 2 is a schematic diagram of a system for a cloud-assisted in-vehicle large-scale language model in at least one embodiment of the present disclosure. The system includes a vehicle 200, a cloud network 209, a first LLM 210, a pre-edit prompt 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, unless otherwise specified, 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.

[0036] In the system of Figure 2, the third LLM 216 is included in the vehicle 200. In at least one embodiment, the third LLM 216 includes a software module installed in the computing system of the vehicle 200. In at least one embodiment, the vehicle 200 is configured to receive the second LLM logit 224 from the cloud network 209, but does not need to be configured to receive the third LLM logit 226 from the cloud network 209. In at least one embodiment, the vehicle 200 requires more computing resources than the vehicle 100 in Figure 1, but the communication required by the vehicle 200 to and from the cloud network 209 is less than the communication required by the vehicle 100 to and from the cloud network 109 in Figure 1.

[0037] Figure 3 is a schematic diagram of a system for a cloud-assisted in-vehicle large-scale language model in at least one embodiment of the present disclosure. The system includes a vehicle 300, a cloud network 309, a first LLM 310, a pre-edit prompt 320, personal information 312, a first LLM logit 321, a second LLM 314, a post-edit prompt 323, a second LLM logit 324, a third LLM 316, a third LLM logit 326, a logit difference 327, and a 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 each of 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] In the system shown in Figure 3, the logit difference 327 and the logit sum 328 are calculated by the cloud network 309. In at least one embodiment, the vehicle 300 is configured to transmit the first LLM logit 321 to the cloud network 309. In at least one embodiment, the cloud network 309 is configured to receive the first LLM logit 321 from the vehicle 300. In at least one embodiment, the cloud network 309 is configured to calculate the logit difference 327 by comparing the first LLM logit 321 with the third LLM logit 326. In at least one embodiment, the cloud network 309 is configured to calculate the logit sum 328 by adding the logit difference 327 to the second LLM logit 324. In at least one embodiment, the cloud network 309 is configured to transmit the logit sum 328 or an output calculated therefrom to the vehicle 300. In at least one embodiment, the computing resources required by vehicle 300 are less than those required by vehicle 100 in Figure 1, and the communication required by vehicle 300 to and from the cloud network 309 is less than the communication required by vehicle 200 to and from the cloud network 209 in Figure 2.

[0040] Figure 4 is a schematic diagram of a system for a cloud-assisted in-vehicle large-scale language model in at least one embodiment of the present disclosure. The system includes a vehicle 400, a cloud network 409, a first LLM 410, a pre-edit prompt 420, personal information 412, a first LLM logit 421, a second LLM 414, a post-edit prompt 423, a second LLM logit 424, a third LLM 416, a third LLM logit 426, a logit difference 427, and a 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, unless otherwise specified, substantially similar in structure and function to each of 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.

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

[0043] In at least one embodiment other than the one shown in Figure 1-4, the configuration and functions may be distributed between the vehicle and the cloud network in different ways so that the pre-edit prompt is not sent from the vehicle and the second LLM logit is processed at least partially using the computing resources of the cloud network.

[0044] Figure 5 is an operational flowchart for a cloud-assisted in-vehicle large-scale language model in at least one embodiment of the present disclosure. In at least one embodiment, the operational flowchart illustrates how to use the cloud-assisted in-vehicle large-scale language model according to at least one embodiment of the present disclosure. In at least one embodiment, the method is performed by a server controller, such as the controller 1002 of the server 1000 shown in Figure 10, which will be described later.

[0045] In S530, the controller inputs a pre-edit prompt to the first LLM. In at least one embodiment, the controller inputs the pre-edit prompt to a first large-scale language model (LLM) installed in the vehicle. In at least one embodiment, the pre-edit prompt contains personal information. In at least one embodiment, the controller obtains a logit output from the first LLM. In at least one embodiment, the logit is used in subsequent operations. In at least one embodiment, the controller executes the operation flow shown in Figure 6, which is described later.

[0046] In S532, the controller adjusts the logit. In at least one embodiment, the controller adjusts the logit output from the first LLM based on other logits from a similar LLM. In at least one embodiment, the controller adjusts the logit based on other logits from a similar cloud-based LLM. In at least one embodiment, the controller calculates the adjusted logit. In at least one embodiment, the controller performs the operation flow shown in Figure 8, which is described later.

[0047] In S533, the controller obtains an iterative output from the adjustment logit. In at least one embodiment, the controller converts the adjustment logit into an output that can be interpreted by a human or computer. In at least one embodiment, the controller obtains the next word in the word sequence.

[0048] In S534, the controller accumulates the iteration outputs. In at least one embodiment, the controller accumulates the iteration outputs from the current iteration to the iteration outputs of previous iterations. In at least one embodiment, the controller accumulates the word sequence following the pre-edit prompt.

[0049] In S536, the controller determines whether the output is complete. If the output is not complete, the operation flow proceeds to S538 to concatenate the iterative output to the pre-edit prompt, and then returns to prompt input in S530. If the output is complete, the operation flow proceeds to S539 to provide or use the output. In at least one embodiment, the controller determines whether to proceed to a further iteration by determining whether the probability of the word sequence being complete is greater than the probability of other words following it.

[0050] In S538, the controller concatenates the iterative output to the pre-edit prompt. In at least one embodiment, the controller concatenates the word sequence to the pre-edit prompt. In at least one embodiment, the concatenation result is input to the first LLM as the updated pre-edit prompt in the next iteration of S530.

[0051] In S539, the controller provides or uses an output. In at least one embodiment, the controller provides an accumulated output to the occupants of the vehicle or uses the output to control the vehicle. In at least one embodiment, the controller displays the accumulated output through the vehicle's display, plays the accumulated output through the vehicle's speakers, etc. In at least one embodiment, the controller uses the accumulated output to control the vehicle or its components, or passes the accumulated output to other components of the vehicle for the control of the vehicle or its components.

[0052] Figure 6 is an operation flowchart for preparing a pre-edit prompt in at least one embodiment of the present disclosure. In at least one embodiment, the operation flowchart illustrates a method for preparing a pre-edit prompt according to at least one embodiment of the present disclosure. In at least one embodiment, the method is performed by a server controller, such as the controller 1002 of the server 1000 shown in Figure 10, which will be described later.

[0053] In S640, the controller or a part thereof detects a raw prompt. In at least one embodiment, the controller identifies and captures a raw prompt, such as an audio sample or an image of an occupant. In at least one embodiment, the controller detects a raw prompt through a microphone, an image sensor, or the like. In at least one embodiment, the controller operates and waits to receive a raw prompt.

[0054] In S642, the controller or a portion thereof computes feature vectors. In at least one embodiment, the controller processes raw prompts to compute one or more feature vectors. In at least one embodiment, the controller extracts embedded information and associates it with personal information. In at least one embodiment, the controller transforms the raw prompts into a form that can be analyzed and processed.

[0055] In S644, the controller or a portion thereof retrieves personal information. In at least one embodiment, the controller analyzes feature vectors to detect personal information about occupants associated with raw prompts. In at least one embodiment, the controller determines which personal information is associated by comparing the feature vectors calculated from the raw prompts with the feature vectors of the personal information. In at least one embodiment, the controller determines associated personal information by the proximity of the feature vectors of the personal information in the feature space to the feature vectors calculated from the raw prompts. In at least one embodiment, the controller calculates the distance between feature vectors in the feature space by applying a distance formula to identify personal information corresponding to the feature vector closest to the feature vector calculated from the raw prompts. In at least one embodiment, the controller uses Retrieval Augmented Generation (RAG) technology.

[0056] In S646, the controller or a portion thereof adds content related to the raw prompt. In at least one embodiment, the controller adds content related to the raw prompt based on identified personal information. In at least one embodiment, as a result of this operation, personal information related to the raw prompt is added. In at least one embodiment, the controller improves the efficiency and accuracy of the LLM output in response to the prompt by performing this process.

[0057] In S648, the controller or a portion thereof places the added prompt into a frame. In at least one embodiment, the controller configures the structure of the added prompt into a format suitable for input to the LLM. In at least one embodiment, the added and framed prompt is used as a pre-edited prompt in a cloud-assisted in-vehicle large-scale language model. In at least one embodiment, the controller uses a processing prompt frame. In at least one embodiment, the controller explores the context in the LLM through the processing prompt frame. In at least one embodiment, the controller uses a prompt frame to inform the LLM that the user is driving the vehicle.

[0058] Figure 7 is a schematic diagram of a vehicle in at least one embodiment 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-edited prompt 720. In at least one embodiment, the vehicle 700 is configured to use RAG technology.

[0059] Vehicle 700, LLM 1 710, personal information 712, and pre-edit prompt 720 are substantially identical in structure and function to Vehicle 100, LLM 1 110, personal information 112, and pre-edit prompt 120, respectively, unless otherwise specified.

[0060] The raw prompt 750 is the first input detected from the user. In at least one embodiment, the raw prompt 750 is a string, a voice command, etc. In at least one embodiment, 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 a raw prompt. In at least one embodiment, the search query generator 718 is configured to generate a query prompt 752 based on a raw prompt 750. In at least one embodiment, 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 one embodiment, the search query generator 718 is a software module, algorithm, etc. In at least one embodiment, the search query generator 718 is configured to receive the search query 754 from the first LLM 710. In at least one embodiment, the search query generator 718 is configured to pass the search query 754 to the feature vector computer 719.

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

[0063] The search query 754 is the output of the first LLM 710 to assist in the identification of personal information. In at least one embodiment, the search query 754 is configured to be coupled with the raw prompt 750.

[0064] The feature vector calculator 719 is configured to compute a feature vector based on a search query. In at least one embodiment, the feature vector calculator 719 is configured to compute a feature vector based on a search query 754 and a raw prompt 750. In at least one embodiment, the feature vector calculator 719 is configured to compute a feature vector 756 by applying a machine learning model.

[0065] Feature vector 756 is used to identify relevant content from personal information 712. In at least one embodiment, feature vector 756 is a numerical vector. In at least one embodiment, feature vector 756 includes embedding information to associate with personal information. In at least one embodiment, feature vector 756 represents the position of search query 754 in feature space.

[0066] Related content 758 represents personal information related to the raw prompt 750. In at least one embodiment, related content 758 is combined with the raw prompt 750 to form the pre-edited prompt 720. In at least one embodiment, related content 758 is a string, multimedia content, etc.

[0067] Figure 8 is an operational flowchart for logit adjustment in at least one embodiment of the present disclosure. In at least one embodiment, the operational flowchart illustrates a method for logit adjustment according to at least one embodiment of the present disclosure. In at least one embodiment, the method is performed by a server controller, such as the controller 1002 of server 1000 shown in Figure 10, which will be described later.

[0068] In the S860, the controller or a portion thereof edits personal information from the pre-edit prompt. In at least one embodiment, the controller enhances privacy by deleting personal information from the pre-edit prompt. In at least one embodiment, the controller executes the operation flow shown in Figure 9, which is described later.

[0069] In S862, the controller or a portion thereof passes the post-edit prompt to a second LLM. In at least one embodiment, the controller passes the post-edit prompt to a second Large Language Model (LLM) located on a cloud network. In at least one embodiment, the controller causes a vehicle transmitter to send the post-edit prompt over a network such as a Wi-Fi network or a wireless access network.

[0070] In S864, the controller or a portion thereof receives the logit output from the second LLM. In at least one embodiment, the controller receives the logit output from the second LLM from the cloud network. In at least one embodiment, the controller retrieves the logit output from the second LLM. In at least one embodiment, the controller instructs the cloud network to transmit the logit output from the second LLM.

[0071] In S865, the controller or a portion thereof inputs the post-edit prompt to the third LLM. In at least one embodiment, the controller inputs the post-edit prompt to the third LLM operating in the vehicle. In at least one embodiment, the controller ensures that the post-edit prompt is input to the third LLM operating in the cloud network.

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

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

[0074] Figure 9 is an operation flowchart for prompt editing in at least one embodiment of the present disclosure. In at least one embodiment, the operation flowchart illustrates a method for prompt editing according to at least one embodiment of the present disclosure. In at least one embodiment, the method is executed by a server controller, such as the controller 1002 of the server 1000 shown in Figure 10, which will be described later.

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

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

[0077] In S974, the controller or a part thereof applies a method for deleting the selected personal information. In at least one embodiment, the controller applies the selected deletion method to the pre-edit prompt. In at least one embodiment, this action effectively edits the identified personal information. In at least one embodiment, this action does not result in any action other than the prompt editing action. In at least one embodiment, this action is performed after the deletion method has been selected. In at least one embodiment, the result of this action is a post-edit prompt with the personal information deleted. In at least one embodiment, this action creates a version of the prompt that respects user privacy.

[0078] In S976, the controller or a part of it determines whether all personal information has been deleted. If it is determined that not all personal information has been completely deleted, the operation flow returns to the detection of personal information in S970. If it is determined that all personal information has been deleted, the operation flow terminates. In at least one embodiment, the controller verifies the post-edit prompt to confirm that all personal information has been properly deleted.

[0079] Figure 10 is a schematic diagram of a hardware configuration for using a cloud-assisted in-vehicle large-scale language model in at least one embodiment of the present disclosure.

[0080] A typical hardware configuration includes a server 1000 that interacts with input devices 1007 directly or through a cloud network 1009. In at least one embodiment, the input device 1007 is a touchscreen, microphone, camera, or other device configured to detect input such as tactile, auditory, or visual. In at least one embodiment, the cloud network 1009 is a wide-area network such as the Internet. In at least one embodiment, the server 1000 is a computer or other computing device that receives input or commands from the input device 1007. In at least one embodiment, the server 1000 is integrated with the input device 1007. In at least one embodiment, the server 1000 is a computer system that executes computer-readable instructions to perform operations for the use of a 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 one embodiment, the controller 1002 includes a processor or programmable circuitry that performs operations according to instructions by executing instructions. In at least one embodiment, the controller 1002 includes analog or digital programmable circuitry, or a combination thereof. In at least one embodiment, the controller 1002 includes physically isolated storage or circuitry that interacts through communication. In at least one embodiment, the storage 1004 includes a non-temporary computer-readable medium capable of storing executable and non-executable data accessed by the controller 1002 during instruction execution. In at least one embodiment, the communication interface 1008 sends and receives data to and from the cloud network 1009. In at least one embodiment, the input / output interface 1006 connects various input and output units, such as input device 1007, via parallel ports, serial ports, keyboard ports, mouse ports, monitor ports, etc., and accepts commands and presents information. In one embodiment, the storage 1004 is located outside the 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] The input unit 1080 is a circuit or instruction of the controller 1002 configured to input prompts to the LLM. In at least one embodiment, the input unit 1080 is configured to input a pre-edit prompt to a first LLM implemented in the vehicle, a post-edit prompt to a second LLM implemented in the cloud network, and a post-edit prompt to a third LLM implemented in either the vehicle or the cloud network. In at least one embodiment, the input unit 1080 uses information in the storage 1004, such as a large language model 1092. In at least one embodiment, the input unit 1080 includes a sub-component for performing additional functions, such as those described in the flowchart above. In at least one embodiment, such a sub-component is referred to by a name related to the corresponding function.

[0084] The adjustment unit 1082 is a circuit or instruction of the controller 1002 configured to adjust the logit. In at least one embodiment, the adjustment unit 1082 calculates the adjusted logit based on the logits output from the first LLM, the second LLM, and the third LLM. In at least one embodiment, the adjustment unit 1082 uses information in the storage 1004, such as the adjustment parameter 1094. In at least one embodiment, the adjustment unit 1082 includes a sub-component for performing additional functions, such as those described in the flowchart above. In at least one embodiment, such a sub-component is referred to by a name related to the corresponding function.

[0085] The editing unit 1084 is a circuit or instruction of the controller 1002 configured to edit the pre-edit prompt. In at least one embodiment, the editing unit 1084 is configured to edit the pre-edit prompt to remove personal information and create a post-edit prompt. In at least one embodiment, the editing unit 1084 uses information in storage 1004, such as personal information 1090 and editing parameters 1096. In at least one embodiment, the editing unit 1084 includes a sub-unit for performing additional functions as described in the flowchart above. In at least one embodiment, such a sub-unit is referred to by a name related to the corresponding function.

[0086] The calculation unit 1086 is a circuit or instruction of the controller 1002 configured to calculate the difference between logits. In at least one embodiment, the calculation unit 1086 is configured to calculate the difference between the logit of the first LLM and the logit of the third LLM. In at least one embodiment, the calculation unit 1086 uses information in the storage 1004, such as the adjustment parameter 1094. In at least one embodiment, the calculation unit 1086 includes a sub-part for performing additional functions, such as those described in the flowchart above. In at least one embodiment, such a sub-part is referred to by a name related to the corresponding function.

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

[0088] In at least one embodiment, the device is a computer, and a program installed on the computer causes the computer to function as the device of the embodiment described herein, or to perform operations associated with the device of the embodiment described herein. In at least one embodiment, such a program is executed on a processor and causes the computer to perform predetermined operations associated with some or all of the blocks in the flowchart and block diagram described herein.

[0089] At least some embodiments are described with reference to flowcharts and block diagrams having (1) steps of how an operation is performed, or (2) blocks representing parts of a controller responsible for performing the operation. In at least some embodiments, certain steps and parts are implemented by dedicated circuits, programmable circuits with computer-readable instructions stored on a computer-readable medium, and / or processors with computer-readable instructions stored on a computer-readable medium. In at least some embodiments, dedicated circuits include digital and / or analog hardware circuits, integrated circuits (ICs), and / or discrete circuits. In at least some embodiments, programmable circuits include reconfigurable hardware circuits such as FPGAs (Field-Programmable Gate Arrays), PLAs (Programmable Logic Arrays), etc., which include logical AND, OR, XOR, NAND, NOR, and other logic operations, flip-flops, registers, memory elements, etc.

[0090] In at least one embodiment, a computer-readable storage medium is a tangible device capable of storing and storing instructions used by an instruction execution device. In one embodiment, the computer-readable storage medium includes, but is not limited to, electronic storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or appropriate combinations thereof. A non-exclusive list of more specific examples of computer-readable storage mediums includes: portable computer diskettes, hard disks, RAM (Random Access Memory), ROM (Read-Only Memory), EPROM (Erasable Programmable Read-Only Memory; flash memory), SRAM (Static Random Access Memory), CD-ROM (Compact Disc Read-Only Memory), DVD (Digital Versatile Disk), Memory Stick®, floppy disks, mechanical encoding devices such as punch cards or grooved raised structures on which instructions are stored, and appropriate combinations thereof. The computer-readable media referred to herein should not be interpreted as transient signals themselves, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves transmitted through waveguides or other transmission media (e.g., light pulses through fiber optic cables), or electrical signals transmitted through wires.

[0091] In at least one embodiment, the computer-readable program instructions described herein are downloadable from a computer-readable storage medium to individual computing / processing devices, or to an external computer or external storage device via a network, such as the Internet, a local area network, a wide area network, and / or a wireless network. In at least one embodiment, the network includes copper transmission cables, optical transmission cables, wireless transmissions, routers, firewalls, switches, gateway computers, and / or edge servers. In at least one embodiment, a network adapter card or network interface of each computing / processing device receives computer-readable program instructions from the network and transfers the computer-readable program instructions for storage on a computer-readable storage medium within the individual computing / processing device.

[0092] In at least one embodiment, the computer-readable program instructions for performing the aforementioned operations are assembler instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, state setting data, or other source code or object code written in a combination of programming languages, including one or more object-oriented programming languages ​​such as Smalltalk, C++, etc., and conventional procedural programming languages ​​such as the C programming language or a similar programming language. In at least one embodiment, the computer-readable program instructions are executed 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 one embodiment, 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) or wide area network (WAN), or a connection to an external computer is formed (for example, via the Internet using an Internet connection provider). In at least one embodiment, an electronic circuit including, for example, a programmable circuit, an FPGA (Field-Programmable Gate Array), or a PLA (Programmable Logic Array) is implemented by executing computer-readable program instructions using state information of computer-readable program instructions to personalize the electronic circuit and carry out the embodiments of the present invention.

[0093] While embodiments of the present invention have been described so far, the technical scope of the claims is not limited to the embodiments described above. Those skilled in the art will understand that various modifications and improvements are possible to the embodiments described above. Furthermore, those skilled in the art will understand from the claims that such modified or improved embodiments are included within the technical scope of the invention.

[0094] The operations, procedures, steps, and stages of each process performed in the apparatus, system, program, and method shown in the claims, embodiments, or drawings may be performed in any order, as long as the order is not indicated by "before," "after," etc., and the output of a preceding process is not used in a subsequent process. Even if a process flow is described in the claims, embodiments, or drawings using phrases such as "first" or "next," such description does not necessarily mean that the processes must be performed in the order described.

[0095] In at least one embodiment, the use of a cloud-assisted in-vehicle large-scale language model is performed by inputting an unedited prompt into a first large-scale language model (LLM) implemented in the vehicle, the unedited prompt containing personal information related to the vehicle's occupants, sending a post-edited prompt to a second LLM implemented in the cloud, the post-edited prompt representing the unedited prompt without personal information, receiving the logit from the second LLM, inputting the post-edited prompt into a third LLM implemented in the cloud or in the vehicle, the third LLM containing the same set of parameters as the first LLM, calculating the difference between the logit from the first LLM and the logit from the third LLM, calculating the sum of that difference and the logit from the second LLM to obtain an adjusted logit, and obtaining an output from the adjusted logit.

[0096] In at least one embodiment, the number of parameters in the second LLM is greater than the number of parameters in the first LLM and the third LLM, respectively. In at least one embodiment, the first LLM is trained using personal information, while the second and third LLMs are trained without personal information. In at least one embodiment, the use of a cloud-assisted in-vehicle large-scale language model includes providing an output to the occupant or using the output to control the vehicle. In at least one embodiment, the use of a cloud-assisted in-vehicle large-scale language model further includes detecting a raw prompt, the raw prompt including at least one of the occupant's voice sample or image. In at least one embodiment, the use of a cloud-assisted in-vehicle large-scale language model further includes adding predetermined information to the raw prompt. In at least one embodiment, the predetermined information includes processed prompt information. In at least one embodiment, the use of a cloud-assisted in-vehicle large-scale language model further includes calculating additional occupant information related to the raw prompt. In at least one embodiment, the unedited prompt includes additional occupant information. In at least one embodiment, the third LLM includes the same parameter values ​​as the first LLM. In at least one embodiment, the third LLM is the first LLM. In at least one embodiment, obtaining an output involves adding up multiple iterative outputs, each iterative output corresponding to iterating and obtaining a reconciled logit in multiple iterative sequences, and the pre-edit prompt in each iteration includes each iterative output corresponding to the preceding iteration in multiple iterative sequences. In at least one embodiment, the use of a cloud-assisted in-vehicle large-scale language model further includes editing personal information from the pre-edit prompt to create a post-edit prompt. In at least one embodiment, editing includes at least one of speech-to-text conversion, image data-to-object data conversion, location generalization, filtering of raw prompts, or randomization of raw prompts. In at least one embodiment, each logit has the same dimensions.

[0097] In at least one embodiment, the use of a cloud-assisted in-vehicle large-scale language model is performed by a device comprising a processor that executes instructions according to the aforementioned operation, or a controller that includes circuits configured to perform the aforementioned operation.

[0098] Having outlined the features of several embodiments to this extent, those skilled in the art will better understand the aspects of this disclosure. Those skilled in the art will recognize that this disclosure can be readily used as the basis for designs or modifications of other processes and structures to achieve the same objectives and / or similar effects as the embodiments described herein. Furthermore, those skilled in the art will realize that such equivalents do not deviate from the spirit and scope of this disclosure, and that various modifications, substitutions, and adjustments described herein are possible without deviating from the spirit and scope of this disclosure.

Claims

1. A pre-edit prompt is entered into the first LLM (Large-Scale Language Model) implemented in the vehicle, and the pre-edit prompt includes personal information related to the occupants of the vehicle. A post-edit prompt is sent to the second LLM implemented in the cloud, and the post-edit prompt represents the pre-edit prompt without personal information. Upon receiving the logit of the second LLM, The edited prompt is entered into the third LLM implemented in the vehicle or the cloud, and the third LLM includes the same parameter set as the first LLM. The difference between the logit of the first LLM and the logit of the third LLM is calculated. The sum of the above difference and the logit of the second LLM is calculated to obtain the adjusted logit. The output is obtained from the aforementioned adjustment logit. A computer program that causes one or more processors to perform an action that includes the following.

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

3. The first LLM is trained using the aforementioned personal information, The second LLM and the third LLM are learned without using the personal information. The computer program according to claim 1 or 2.

4. The computer program according to claim 1 or 2, wherein the operation further includes providing the output to the occupant or using the output to control the vehicle.

5. The computer program according to claim 1 or 2, wherein the operation further includes detecting a raw prompt, the raw prompt comprising at least one of the occupant's voice sample or image.

6. The computer program according to claim 5, wherein the operation further includes adding predetermined information to the raw prompt.

7. The computer program according to claim 6, wherein the predetermined information includes processing prompt information.

8. The computer program according to claim 5, wherein the operation further includes calculating additional occupant information related to the raw prompt.

9. The computer program according to claim 8, wherein the pre-editing prompt includes the additional occupant information.

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

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

12. Obtaining the aforementioned output involves adding up multiple iterative outputs, Each iteration output corresponds to repeatedly obtaining the adjustment logit in multiple iteration sequences. The pre-edit prompt in each iteration includes the iteration output corresponding to the preceding iteration in the plurality of iteration sequences. The computer program according to claim 1 or 2.

13. The computer program according to claim 5, wherein the operation further includes editing the personal information from the pre-edit prompt to create the post-edit prompt.

14. The computer program according to claim 13, wherein the editing includes at least one of converting speech to text, converting image data to object data, generalizing location, filtering the raw prompts, or randomizing the raw prompts.

15. The computer program according to claim 1 or 2, wherein each logit has the same dimension.

16. A method that is executed by one or more processors, A pre-edit prompt is entered into the first LLM (Large-Scale Language Model) implemented in the vehicle, and the pre-edit prompt includes personal information related to the occupants of the vehicle. A post-edit prompt is sent to the second LLM implemented in the cloud, and the post-edit prompt represents the pre-edit prompt without personal information. Upon receiving the logit of the second LLM, The edited prompt is entered into the third LLM implemented in the vehicle or the cloud, and the third LLM includes the same parameter set as the first LLM. The difference between the logit of the first LLM and the logit of the third LLM is calculated. The sum of the above difference and the logit of the second LLM is calculated to obtain the adjusted logit. The output is obtained from the aforementioned adjustment logit. A method that includes the act of doing so.

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

18. The first LLM is trained using the aforementioned personal information, The second LLM and the third LLM are learned without using the personal information. The method according to claim 16 or 17.

19. The method according to claim 16 or 17, further comprising providing the output to the occupant or using the output to control the vehicle.

20. A pre-edit prompt is entered into the first LLM (Large-Scale Language Model) implemented in the vehicle, and the pre-edit prompt includes personal information related to the occupants of the vehicle. A post-edit prompt is sent to the second LLM implemented in the cloud, and the post-edit prompt represents the pre-edit prompt without personal information. Upon receiving the logit of the second LLM, The edited prompt is entered into the third LLM implemented in the vehicle or the cloud, and the third LLM includes the same parameter set as the first LLM. The difference between the logit of the first LLM and the logit of the third LLM is calculated. The sum of the above difference and the logit of the second LLM is calculated to obtain the adjusted logit. The output is obtained from the aforementioned adjustment logit. A device comprising a controller including a circuit configured to perform an operation that includes the following.

Images