Voice interaction model optimization method, interaction method, vehicle and storage medium

Abstract

The invention relates to a voice interaction model optimization method, an interaction method, a vehicle and a storage medium. The method comprises the following steps: receiving a first voice interaction instruction; constructing an initial thinking chain corresponding to the first voice interaction instruction; performing self-checking on the initial thinking chain, and generating a corresponding target thinking chain based on a self-checking result; the self-checking is used for identifying whether an error exists in the initial thinking chain or not; and optimizing a pre-configured voice interaction model according to the target thinking chain. According to the scheme provided by the invention, the voice interaction accuracy between the user and the vehicle-mounted terminal can be effectively improved.

Description

Technical Field

[0001] This application relates to the field of voice interaction technology, and in particular to optimization methods for voice interaction models, interaction methods, vehicles, and storage media. Background Technology

[0002] With the widespread application of intelligent interaction technology in various terminal devices, intelligent voice-driven methods are gradually becoming an important interaction means. Generally speaking, the intelligent voice-driven process involves ASR (Automatic Speech Recognition) and RAG (Retrieval-Augmented Generation).

[0003] Among them, ASR is responsible for converting acoustic signals into computationally processable text, while RAG is responsible for searching similar APIs (Application Programming Interfaces) in a massive database, concatenating the ASR recognition results with similar APIs to obtain new prompt words, so that NLU (Natural Language Understanding) can combine the new prompt words to generate executable instructions that can accurately express the user's intentions.

[0004] However, current large models rely on a human-defined vehicle knowledge system, the coverage of which is difficult to fully reflect the real vehicle interaction environment. This makes the large models susceptible to the influence of pre-existing knowledge during actual reasoning, thereby reducing the accuracy of voice interaction between users and in-vehicle terminals. Summary of the Invention

[0005] To address or partially address the problems existing in related technologies, this application provides an optimization method for a voice interaction model, an interaction method, a vehicle, and a storage medium, which can effectively improve the accuracy of voice interaction between users and in-vehicle terminals.

[0006] The first aspect of this application provides a method for optimizing a voice interaction model, comprising: Receive the first voice interaction command; Construct the initial thought chain corresponding to the first voice interaction command; The initial thought chain is self-verified, and a corresponding target thought chain is generated based on the self-verification result; the self-verification is used to identify whether there are errors in the initial thought chain. The pre-configured voice interaction model is optimized based on the target thought chain.

[0007] In one instance, the step of self-verifying the initial thought chain and generating a corresponding target thought chain based on the self-verification result includes: Extract the real labels from the first voice interaction command; The initial thought chain is sampled multiple times to obtain predicted labels; The self-verification result of the initial thought chain is determined based on the comparison relationship between the real label and the predicted label. Based on the self-verification results, the initial thought chain is subjected to hierarchical reasoning to generate the target thought chain.

[0008] In one instance, the step of performing hierarchical reasoning on the initial thought chain according to the self-verification result to generate the target thought chain includes: When there is at least one sampling operation whose predicted label is the same as the real label corresponding to the first voice interaction instruction, obtain a reference thought chain example of the first voice interaction instruction. Based on the reference thought chain example, the output format of the initial thought chain is adjusted to generate the target thought chain.

[0009] In one instance, the step of performing hierarchical reasoning on the initial thought chain according to the self-verification result to generate the target thought chain includes: When the predicted label corresponding to a continuous sampling operation is different from the real label corresponding to the first voice interaction command, the vehicle knowledge associated with the first voice interaction command is determined. Based on the vehicle knowledge, contradiction detection is performed on the first voice interaction command and the initial thought chain to obtain self-correction information; Based on the self-correcting information, the target thought chain is regenerated.

[0010] In one instance, optimizing the pre-configured voice interaction model according to the target thought chain includes: The first voice interaction command and the initial thought chain are combined into a first training sample; The first voice interaction command, the self-correction information, and the target thought chain are combined into a second training sample. The pre-configured voice interaction model is optimized based on the first training sample and the second training sample.

[0011] In one instance, optimizing the pre-configured voice interaction model based on the first training sample and the second training sample includes: During the process of training the voice interaction model using the first training sample, the loss value of the first training sample is not calculated. During the process of training the voice interaction model using the second training sample, the loss value of the second training sample is calculated, and the model parameters of the voice interaction model are updated based on the loss value until the output result of the voice interaction model meets the preset conditions.

[0012] In one instance, constructing the initial thought chain corresponding to the first voice interaction command includes: Parse the first voice interaction command to obtain the voice interaction text and the vehicle knowledge associated with the voice interaction text; Retrieve a list of application programming interfaces (APIs) corresponding to the voice interaction text; the list of APIs includes several APIs and interface parameters corresponding to each API. The vehicle knowledge, the application programming interface, and the interface parameters are filled into a preset prompt word template to obtain the initial prompt word; The initial thought chain is generated by performing thought chain reasoning on the initial prompt words.

[0013] A second aspect of this application provides a voice interaction method based on a voice interaction model, comprising: Real-time acquisition of second voice interaction commands; The second voice interaction command is input into the optimized voice interaction model to generate the corresponding vehicle control task.

[0014] A third aspect of this application provides a vehicle, comprising: Processor; and A memory that stores executable code, which, when executed by the processor, causes the processor to perform the method described above.

[0015] A fourth aspect of this application provides a computer-readable storage medium having executable code stored thereon, which, when executed by a processor of an electronic device, causes the processor to perform the method described above.

[0016] The fifth aspect of this application provides a computer program product comprising computer instructions that, when executed by a processor, implement the method described above.

[0017] The technical solution provided in this application may include the following beneficial results: In this application, a first voice interaction command is received; an initial thought chain corresponding to the first voice interaction command is constructed; the initial thought chain is self-verified, and a corresponding target thought chain is generated based on the self-verification result; the self-verification is used to identify whether there are errors in the initial thought chain; and a pre-configured voice interaction model is optimized according to the target thought chain.

[0018] Compared to related technologies, the technical solution of this application makes the abstract reasoning process explicit by constructing an initial thought chain corresponding to the first voice interaction command. This explicit reasoning structure helps to fully understand the user's intent and improves the optimization effect of the voice interaction model. Furthermore, a self-reflection mechanism is introduced into the model reasoning process, enabling the model to autonomously identify reasoning deviations and generate a more reasonable target thought chain based on the self-verification results of the initial thought chain. This target thought chain is then used to specifically optimize the voice interaction model, significantly improving the reasoning accuracy of the voice interaction model.

[0019] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0020] The above and other objects, features and advantages of this application will become more apparent from the more detailed description of exemplary embodiments thereof in conjunction with the accompanying drawings, wherein the same reference numerals generally represent the same components in the exemplary embodiments thereof.

[0021] Figure 1 This is a flowchart illustrating an optimization method for a voice interaction model according to an embodiment of this application; Figure 2 This is another flowchart illustrating an optimization method for a voice interaction model as shown in an embodiment of this application; Figure 3 This is a schematic diagram of the model self-reflection process shown in the embodiments of this application; Figure 4 This is a schematic diagram of the voice interaction model training process shown in the embodiments of this application; Figure 5 This application illustrates a voice interaction method based on a voice interaction model. Figure 6 This is a schematic diagram illustrating the process of online voice interaction based on complex COT as shown in the embodiments of this application; Figure 7 This is a schematic diagram of the vehicle structure shown in the embodiments of this application. Detailed Implementation

[0022] Embodiments of this application will now be described in more detail with reference to the accompanying drawings. While embodiments of this application are shown in the drawings, it should be understood that this application may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to make this application more thorough and complete, and to fully convey the scope of this application to those skilled in the art.

[0023] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used in this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0024] It should be understood that although the terms "first," "second," "third," etc., may be used in this application to describe various information, this information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0025] With the widespread application of intelligent interaction technology in various terminal devices, intelligent voice-driven methods are gradually becoming an important interaction means. Generally speaking, the intelligent voice-driven process involves ASR (Automatic Speech Recognition) and RAG (Retrieval-Augmented Generation).

[0026] Among them, ASR is responsible for converting acoustic signals into computationally processable text, while RAG is responsible for searching similar APIs (Application Programming Interfaces) in a massive database, concatenating the ASR recognition results with similar APIs to obtain new prompt words, so that NLU (Natural Language Understanding) can combine the new prompt words to generate executable instructions that can accurately express the user's intentions.

[0027] However, current large models rely on a human-defined vehicle knowledge system, the coverage of which is difficult to fully reflect the real vehicle interaction environment. This makes the large models susceptible to the influence of pre-existing knowledge during actual reasoning, thereby reducing the accuracy of voice interaction between users and in-vehicle terminals.

[0028] To address the aforementioned issues, this application provides an optimization method for a voice interaction model, which can effectively improve the accuracy of voice interaction between users and in-vehicle terminals.

[0029] The technical solution of this application is applicable to human-vehicle interaction scenarios in the vehicle field. The following explanation uses a scenario where a user is in a smart cockpit and issues voice interaction commands from within the smart cockpit as an example.

[0030] The intelligent cockpit refers to an integrated digital platform that integrates artificial intelligence, communication technology, and electronic control systems into the in-vehicle environment. Compared with traditional cockpits, intelligent cockpits have multiple interaction methods such as voice, touch, visual recognition, and gesture perception, offering a more natural human-machine dialogue and more comprehensive in-vehicle function control capabilities.

[0031] In this application, the data processing center of the intelligent cockpit is the vehicle-mounted terminal, which works in conjunction with other devices to complete vehicle control tasks.

[0032] As an example, a smart cockpit is equipped with a voice interaction device that collects voice commands issued by the user. For instance, multiple microphones are installed in the smart cockpit to collect audio data issued by the user in real time. In actual use, the coverage of voice signal collection can be improved by arranging the microphones in different ways.

[0033] In addition, the smart cockpit is also equipped with actuators to perform vehicle control tasks. For example, the seat assembly is used to perform vehicle control tasks such as seat position adjustment, heating control, and ventilation control; the vehicle air conditioning assembly is used to perform vehicle control tasks such as temperature adjustment, air volume adjustment, and airflow direction switching; and the window assembly is used to perform vehicle control tasks such as window raising and lowering, and sunroof opening / closing.

[0034] In addition, considering the need to provide voice prompts to users, the smart cockpit can also be equipped with audio playback devices. For example, multiple speakers can be installed in the smart cockpit, arranged in a surround or zoned manner, to ensure that users can still clearly hear the voice prompts even when the vehicle is in operation.

[0035] As described above regarding the smart cockpit, the main entity handling data processing within the smart cockpit is actually the in-vehicle terminal. In practical applications, the in-vehicle terminal can flexibly deploy different types of large models according to different data processing tasks.

[0036] As an example, when data processing tasks involve reasoning about voice interaction commands issued by users in a smart cockpit, the in-vehicle terminal can simultaneously deploy a voice interaction model and a large-scale thought chain correction model. Both the voice interaction model and the large-scale thought chain correction model belong to large language models that can understand user intentions and generate natural language text.

[0037] The voice interaction model refers to a large model used to parse the intent of voice interaction commands and generate corresponding vehicle control tasks. This model mainly undertakes online inference, and can quickly complete tasks such as semantic understanding, slot extraction, and control task generation after receiving voice interaction commands issued by the user in real time, thereby meeting the requirements of high real-time performance and high response speed of intelligent cockpits in actual driving scenarios.

[0038] The large-scale thinking chain correction model refers to a large model with stronger reasoning depth and self-correction capabilities. It can perform reasoning analysis, error location and correction prompts on the initial thinking chain of voice interaction commands, and generate target thinking chains based on the correction prompt information, thereby using the target thinking chain to correct the voice interaction model.

[0039] Furthermore, the self-correcting ability of the large-scale thinking chain correction model is built upon the SR-COT (Self-Reflective Chain-of-Thought) strategy. In complex intelligent cockpit scenarios (such as multi-task concurrency and fuzzy commands), traditional COT technology is prone to misleading the model into making erroneous inferences, and such models lack the ability to verify and correct their own reasoning processes.

[0040] This application employs a hierarchical SR-COT strategy to enable the large-scale thought chain correction model to autonomously analyze the cause of errors by combining vehicle knowledge and voice interaction commands, even if an incorrect initial thought chain is generated. Based on this analysis, the model corrects the initial thought chain to obtain a target thought chain that conforms to the actual intelligent cockpit environment. Ultimately, the target thought chain guided by the SR-COT strategy can be used to train and optimize the voice interaction model, improving its inference accuracy and generalization ability.

[0041] Compared to the large model of thought chain correction, the above-mentioned voice interaction model is smaller in scale and has a more compact structure, making it more suitable for performing online inference tasks in in-vehicle terminals, ensuring that the actual voice interaction response speed and accuracy remain at a high level.

[0042] The large-scale thinking chain correction model requires in-depth reasoning analysis and error localization of the initial thinking chain. Its reasoning process is relatively complex and computationally intensive, making it more suitable as an optimization module for generating complex COTs. The complex COTs generated by the large-scale thinking chain correction model not only retain prior knowledge but also incorporate subsequent knowledge such as semantic information, control logic, and device relationships of the smart cockpit. This allows the voice interaction model optimized based on the complex COT to fully understand the complex vehicle interaction environment and thus accurately infer voice interaction commands.

[0043] For example, the voice interaction model is a 7B model (7-billion-parameter model). The 7B model has a relatively moderate number of parameters, enabling it to complete tasks such as voice understanding, real-time response, and short text generation on the device side with low latency, thus ensuring the response speed and system stability required for voice interaction scenarios. The thought chain correction model is a 72B model (72-billion-parameter model), which, with its significantly larger parameter capacity and deeper representation capabilities than the 7B model, can correct and optimize inference paths. Integrating these two models into the in-vehicle terminal significantly improves the accuracy of inference while maintaining a natural and fluent output.

[0044] The technical solutions of the embodiments of this application are described in detail below with reference to the accompanying drawings.

[0045] Figure 1 This is a flowchart illustrating an optimization method for a voice interaction model according to an embodiment of this application.

[0046] See Figure 1 This method is applied to an in-vehicle terminal, which is equipped with a voice interaction model and a large model for correcting thought chains. The method includes at least the following steps.

[0047] Step 101: Receive the first voice interaction command.

[0048] In this embodiment, the thought chain correction model can receive the first voice interaction command initiated by the user in the smart cockpit.

[0049] Optionally, a user refers to a driver or passenger who is inside the smart cockpit environment and initiates interactive commands to the vehicle via voice. For example, a driver in the driver's seat, a passenger in the front passenger seat, or a passenger in the rear seat.

[0050] The first voice interaction command refers to a voice interaction command that has been successfully collected in the past and responded to by the voice interaction model. This first voice interaction command is stored in the historical interaction database. In this application, the first voice interaction command is used as a high-quality sample for optimizing the voice interaction model.

[0051] For example, the first voice interaction command might be: Within the past day, a user in the smart cockpit issued the command "Adjust the driver's seat back," and the in-vehicle terminal successfully interpreted and accurately executed the control operation "Adjust the driver's seat back to the user's preferred angle." Alternatively, the first voice interaction command might be: Within the past week, a user in the smart cockpit issued the command "Turn off the in-vehicle sound," and the in-vehicle terminal successfully interpreted and accurately executed the control operation "Turn off the audio in the smart cockpit."

[0052] At this point, the instruction not only clearly expresses the user's intent but also carries reliable response data, so it can be used to optimize the large model of thought chain correction.

[0053] Step 102: Construct the initial thought chain corresponding to the first voice interaction command.

[0054] In this embodiment, the thought chain correction model analyzes the first voice interaction command based on its original reasoning structure, identifies the user intent, possible interaction objects and interaction operations in the first voice interaction command, and thus obtains the initial thought chain corresponding to the first voice interaction command.

[0055] Optionally, the initial thought chain refers to the reasoning process text generated based on the first voice interaction command and the vehicle knowledge and application programming interface related to that command. The initial thought chain typically reflects the original thought process of the larger thought chain correction model regarding the first voice interaction command before optimization, and therefore may contain defects such as incomplete logic and reasoning biases. These defects are precisely what the subsequent correction process needs to identify and correct.

[0056] As an example, when the first voice interaction command is "lower the right seat cushion of the third row a little", the thought chain correction model may generate the initial thought chain according to the following reasoning structure: 1) Understand the first voice interaction command and clarify that the user's intention is to lower the position of the right seat cushion of the third row; 2) Refer to vehicle knowledge and learn that the seat cushion in the car supports up and down adjustment; 3) Select the appropriate API and determine the interface Seat_Gear_Set; 4) Fill in the API interface parameters: device=seat cushion, direction=lower, set_type=lower, seat_position=right side of the third row.

[0057] Step 103: Perform self-verification on the initial thought chain and generate the corresponding target thought chain based on the self-verification result. The self-verification is used to identify whether there are errors in the initial thought chain.

[0058] In this embodiment, the large-scale thinking chain correction model first self-verifies the initial thinking chain to determine whether self-reflection is necessary. If self-reflection is required, the model outputs a self-verification result after completion, generates a target thinking chain based on the result, and optimizes the pre-configured voice interaction model according to the target thinking chain. If self-reflection is not required, it indicates that the initial thinking chain generated by the large-scale thinking chain correction model can correctly capture the user's intent of the voice interaction command, and the reasoning logic is clear and consistent with prior knowledge. In this case, additional self-reflection will not bring any benefit but may increase computational overhead; therefore, there is no need to generate an additional target thinking chain.

[0059] Self-verification refers to the process by which the large-scale thinking chain correction model combines multiple inference sampling results to identify whether there are errors in the initial thinking chain. In fact, self-verification can be built upon the self-reflection mechanism of the large-scale thinking chain correction model. Self-reflection means that after generating the initial thinking chain, the large-scale thinking chain correction model examines its own reasoning process and autonomously identifies potential logical biases, misunderstandings of user intent, or knowledge conflicts.

[0060] The self-verification result indicates the cause of the error in the initial thought chain. For example, the self-verification result includes two situations: "The reasoning path of the initial thought chain is correct, but the output format is not standardized" and "The reasoning path of the initial thought chain has deviations".

[0061] Based on the self-verification results, the thought chain correction model can generate more reasonable target thought chains that meet the needs of real-world scenarios. It is evident that the core of self-reflection lies in the model autonomously identifying problems and automatically correcting them, without relying on external intervention. Optionally, the target thought chain refers to the reasoning process text generated by the thought chain correction model after self-reflection on the initial thought chain, accurately reflecting the user's true intent and expected reasoning path. The target thought chain can serve as standard data to guide the optimization of the internal reasoning structure of the voice interaction model, thereby improving the voice interaction model's understanding of voice interaction commands and the accuracy of its decision-making.

[0062] Step 104: Optimize the pre-configured voice interaction model based on the target thinking chain.

[0063] In this embodiment, the thought chain correction model can optimize the pre-configured voice interaction model based on the target thought chain, enabling the voice interaction model to learn the reasoning structure and decision-making logic of the target thought chain and improve the performance of the voice interaction model.

[0064] Among them, optimization refers to the large model of thinking chain correction using the target thinking chain as a high-quality supervision signal, and training the current voice interaction model through supervised fine-tuning, so that it gradually learns the reasoning structure and decision-making logic of the target thinking chain.

[0065] The optimized voice interaction model will be deployed back to the in-vehicle terminal to perform online inference tasks, thereby directly outputting highly accurate vehicle control tasks in subsequent real-time voice interaction requests.

[0066] As an example, if the initial thought chain results in one or more of the following: uncertain prediction results, prediction deviation exceeding a preset threshold, or inconsistency with vehicle knowledge, a self-reflection mechanism is automatically triggered. Conversely, if the initial thought chain reasoning result is reliable and consistent with vehicle knowledge, the initial thought chain can be executed directly without further reflection.

[0067] In this embodiment, a first voice interaction command is received; an initial thought chain corresponding to the first voice interaction command is constructed; the initial thought chain is self-verified, and a corresponding target thought chain is generated based on the self-verification result; the self-verification is used to identify whether there are errors in the initial thought chain; and the pre-configured voice interaction model is optimized according to the target thought chain.

[0068] Compared to related technologies, the technical solution of this application makes the abstract reasoning process explicit by constructing an initial thought chain corresponding to the first voice interaction command. This explicit reasoning structure helps to fully understand the user's intent and improves the optimization effect of the voice interaction model. Furthermore, a self-reflection mechanism is introduced into the model reasoning process, enabling the model to autonomously identify reasoning deviations and generate a more reasonable target thought chain based on the self-verification results of the initial thought chain. This target thought chain is then used to specifically optimize the voice interaction model, significantly improving the reasoning accuracy of the voice interaction model.

[0069] Figure 2 This is another flowchart illustrating an optimization method for a voice interaction model as shown in an embodiment of this application. Figure 2 relatively Figure 1 The technical solutions of the embodiments of this application are described in more detail.

[0070] Step 201: Receive the first voice interaction command.

[0071] For complex smart cockpit scenarios, the Mind Chain Correction Model can retrieve the first voice interaction command initiated by the user in the smart cockpit from the historical interaction database. For example, the first voice interaction command might be "lower the right seat cushion in the third row a little".

[0072] Step 202: Construct the initial thought chain corresponding to the first voice interaction command.

[0073] After receiving the first voice interaction command, the large-scale thinking chain correction model triggers a complex COT generation command to obtain the initial thinking chain corresponding to the first voice interaction command.

[0074] It is worth noting that the data basis for optimizing the voice interaction model in this application is the initial thought chain obtained by the thought chain correction model based on the first voice interaction command processed by the voice interaction model. By reflecting on the initial thought chain, the thought chain correction model can generate a more accurate and matching target thought chain, providing high-quality samples for optimizing the voice interaction model.

[0075] The response process for complex COT generation commands includes at least the following steps: parsing the first voice interaction command to obtain the voice interaction text and the vehicle knowledge associated with it; retrieving the application programming interface (API) list corresponding to the voice interaction text, which includes several APIs and their corresponding interface parameters; filling the vehicle knowledge, APIs, and interface parameters into a preset prompt word template to obtain initial prompt words; and performing thought chain reasoning on the initial prompt words to generate an initial thought chain.

[0076] The voice interaction text refers to the text transcribed from the first voice interaction command using voice recognition technology. This text not only contains the original text content of the first voice interaction command, but may also include contextual information, such as the vehicle's interaction environment at the time the first voice interaction command was issued.

[0077] Vehicle knowledge refers to the prior knowledge related to smart cockpit functions that is manually compiled and added during the construction of application programming interfaces (APIs). This knowledge assists the large model in understanding and reasoning about user-issued voice interaction commands. Vehicle knowledge includes, but is not limited to, the parameter types, range values, and interface data of each device in the smart cockpit. For example, vehicle knowledge might include the following: "The seat cushion can be adjusted up and down and forward and backward," "The leg rest can be adjusted as part of the seat cushion," "The lumbar support supports forward and backward and up and down adjustment," and "The vehicle is equipped with three rows of seats."

[0078] The application programming interface (API) list refers to the collection of available APIs in the smart cockpit. Each API is the entry point for the smart cockpit control functions, which is to say, the smart cockpit control functions are abstracted into callable interfaces.

[0079] Interface parameters refer to the call information corresponding to each application interface. For example, interface parameters include the parameter name of the PI, the parameter type (enumeration, string, number), the range of selectable values, etc., thereby constraining the content of the API call and avoiding interface call failure due to incorrect API parameters when generating vehicle control tasks.

[0080] Preset prompt templates refer to pre-defined text templates used to guide the voice interaction model or the large-scale thought chain correction model in generating thought chains. These templates specify the prompt structure, output format requirements, and reasoning steps needed to generate thought chains, thereby ensuring that the generated thought chains meet expectations in terms of logical order, information completeness, and format standardization.

[0081] The initial prompt word refers to the concatenated prompt word obtained by filling the voice interaction text, the retrieved application interface list, the corresponding interface parameters, and relevant vehicle knowledge into the preset prompt word template. The initial prompt word is used to guide the mind chain correction model to generate the initial mind chain.

[0082] As an example, a preset prompt template could be: "You are an in-car voice assistant that interacts with the user based on their commands, the APIs and their parameters provided in the candidate API list, and the vehicle knowledge given." 1. First, you should think based on the user instructions and your vehicle knowledge, and provide your thought process. 2. Based on logical reasoning, determine whether the following APIs can help you complete the task. If no API can help you complete the task, output "No". 3. If a suitable API is available in the list, select one and then fill in the parameters for that API. 4. Note that if a parameter is required, you must fill it in; if it is an enumeration value, you can only choose from the listed candidate values. [Your output format] Logical thinking: 1. Understanding user instructions: 2. Refer to vehicle knowledge: 3. Choose the appropriate API: 4. Populate API parameters Is there a suitable API: Yes or No Selected API: {"API": "api1", "ARGUMENTS": {"arg1": "value1", "arg2": "value2", ... ,"argn": "valuen"}} [Vehicle Knowledge] The lumbar support can be adjusted forward and backward, as well as up and down. Leg rests can be understood as seat cushions. The seat cushion can be adjusted up and down, and also forward and backward. [List of available APIs]

Seat_Value_Set

[0083] ARGUMENTS: "device": (enumerated value, required) Optional value: one of ['backrest', 'seat', 'leg rest', 'headrest', 'cushion', 'lumbar support']. "direction": (enumeration value, required) Optional values: ['up', 'down', 'forward', 'back', 'middle'] "set_type": (enumeration value, required) The value can be one of the following: ['Raise', 'Lower', 'To', 'Before', 'After']. "seat_position": (enumeration value, optional) Optional values ​​include ['driver', 'passenger', 'front row', 'second row left', 'second row right', 'second row', 'all', 'third row left', 'third row right', 'third row'] "set_value": (string, required) This parameter has no limit on the range of values. It can be one of two types: a number (e.g., 1 to 100) or a relative value (e.g., 3%, half, middle, one-third). Both types of values ​​should be output as text in the command; no numerical conversion is required.

Seat_Gear_Set

[0084] ARGUMENTS: "device": (enumerated value, required) Optional value: one of ['backrest', 'seat', 'leg rest', 'headrest', 'cushion', 'lumbar support']. "direction": (enumeration value, required) Optional values: ['up', 'down', 'forward', 'back', 'middle'] "set_type": (enumeration value, required) The value can be one of ['increment', 'decrement', 'before', 'after']. "seat_position": (enumeration value, optional) Optional values ​​include ['driver', 'passenger', 'front row', 'second row left', 'second row right', 'second row', 'all', 'third row left', 'third row right', 'third row']

Seat_Max_Min_Set

[0085] ARGUMENTS: "device": (enumerated value, required) Optional value: one of ['backrest', 'seat', 'leg rest', 'headrest', 'cushion', 'lumbar support']. "direction": (enumeration value, required) Optional values: ['up', 'down', 'forward', 'back', 'middle'] "set_value": (string, required) Optional values ​​['lowest', 'highest', 'first', 'last']. This parameter is one of the following: "seat_position": (enumeration value, optional) Values ​​can be one of ['driver', 'passenger', 'front row', 'second row left', 'second row right', 'second row', 'all', 'third row left', 'third row right', 'third row']. User commands "Adjust the right seat cushion in the third row to a tiny bit lower."

[0086] Thinking chain reasoning refers to the process text that reflects internal reasoning, which is generated step by step by the thinking chain correction model according to the preset reasoning steps, taking the initial prompt words as input. This process includes understanding and rewriting the instructions, making feasibility judgments by referring to vehicle knowledge, selecting the most suitable interface from the application interface list and filling in the parameters one by one, and finally outputting the initial thinking chain.

[0087] For example, the thought chain reasoning process generates the initial thought chain step by step according to the following preset reasoning steps: 1) For incomplete or ambiguous voice interaction commands, the model first parses and rewrites the voice interaction commands to make their semantics clearer and more executable. 2) Combining information such as device functions, parameter ranges, and constraints in the vehicle knowledge base, the model is assisted in judging the feasibility of the command, which is equivalent to injecting prior knowledge into the model. 3) Based on understanding the voice interaction commands, the API that best matches the user's intent is selected from the list of available application programming interfaces. 4) According to the parameter definitions of the selected API and the voice interaction commands, the API parameters are filled in item by item to ensure the completeness and correctness of the interface call.

[0088] As an example, firstly, the thought chain correction model parses the first voice interaction command based on its configured ASR module, outputting the voice interaction text "Turn off the sound in the car." Then, it uses the RAG module to retrieve the top-K similar APIs semantically closest to this voice interaction text, and simultaneously uses the BGE (Big Generalist Embedding) model to retrieve the interface parameters of each API from the API description vector library. Next, it fills the preset prompt word template with vehicle knowledge and the top-K similar APIs and their interface parameters to generate the initial prompt word. This initial prompt word explicitly requires the thought chain correction model to output the thinking logic, the existence of a suitable API, and the selected API and its parameter filling results according to a preset data format. Finally, the thought chain correction model performs thought chain reasoning based on this initial prompt word to generate the initial thought chain.

[0089] Step 203: Perform self-verification on the initial thought chain and generate the corresponding target thought chain based on the self-verification results.

[0090] The large-scale model for correcting thought chains can extract real labels from the first voice interaction command and perform multiple sampling operations on the initial thought chain to obtain predicted labels. Based on the comparison between the predicted labels and the real labels, the self-verification result of the initial thought chain is determined, and hierarchical reasoning is performed on the initial thought chain according to the self-verification result to generate the target thought chain.

[0091] In this context, a label refers to the target value or expected output result associated with a sample. For example, in model optimization, labels can be manually labeled or automatically created identifiers such as categories, values, and attributes.

[0092] Authentic labels refer to the confirmed authentic results for the first voice interaction command, typically expressed as follows:<API,Arguments> The format is represented as follows: it includes the name of the interface to be called and the corresponding interface parameters.

[0093] Predicted labels refer to the natural language understanding results generated each time the initial thought chain is sampled multiple times.<API,Arguments> The prediction label reflects the prediction results given by the large model of thought chain correction under the current inference structure.

[0094] When comparing predicted and real labels, the comparison is not limited to complete lexical equality. Instead, it allows for the determination of the matching relationship by calculating the semantic similarity between the two. In other words, as long as the predicted and real labels are semantically consistent, even if there are synonym substitutions, differences in parameter descriptions, or different word orders, they can still be considered a match.

[0095] Sampling refers to calling the thought chain correction model multiple times (denoted as N times) under the same initial prompt words to obtain N independent inference outputs. Each inference output is then parsed using structured parsing to obtain a predicted label.<API,Arguments> Multiple samplings were used to reflect the stability and consistency of the reasoning of the modified thinking chain model under the current initial thinking chain. If there are large differences between the predicted labels generated by different sampling operations, it indicates that the model reasoning is unstable and needs further correction. If the predicted labels are highly consistent with the true labels, it indicates that the current initial thinking chain is of high quality.

[0096] As an example, the thought chain correction model itself possesses a certain degree of instruction-following ability. By increasing the number of samples and the sampling probability, it may autonomously generate correct inference results. Compared to externally added thought chains, thought chains generated autonomously by the model tend to be closer to its inherent inference structure and are more suitable as high-quality samples for subsequent training. Therefore, this step involves independently calling the current thought chain correction model N times under the same initial prompt words to obtain N inference outputs. The value of N can be flexibly set according to computational resources and accuracy requirements, such as 5 times, 10 times, or 20 times.

[0097] Preferably, in order to achieve a balance between sampling stability and computational overhead, this application sets N to 10.

[0098] refer to Figure 3 , Figure 3 This is a schematic diagram of the model self-reflection process shown in the embodiments of this application. During the 10 sampling cycles, if any predicted label obtained from a sampling matches the true label (denoted as T1), there is no need to rely on a larger model to regenerate the complete inference chain. Instead, the output format of the initial thought chain is corrected to conform to the preset structure and inference paradigm, thus obtaining the target thought chain. If the correct result is not obtained for 10 consecutive sampling cycles (denoted as T0), it indicates that the current thought chain correction model lacks sufficient inference ability on this type of instruction and is unable to obtain an accurate thought chain through self-sampling. In this case, this application provides the correct natural language understanding result to the thought chain correction model as a prompt, enabling the model to perform hierarchical inference based on the correct natural language understanding result after obtaining the initial thought chain, thereby generating a more accurate target thought chain.

[0099] Furthermore, in case T1, even if the initial thought chain is semantically correct, there may still be issues such as non-standard expression, inconsistent field formats, and missing required parameters.

[0100] In this case, the self-verification result of the initial thought chain is that there is at least one sampling operation whose predicted label is the same as the real label corresponding to the first voice interaction command. In other words, the reasoning path of the initial thought chain is correct, but the output format is not standardized.

[0101] At this point, it can be assumed that the current large-scale model for correcting thought chain has the ability to reason correctly, but the output results are unstable due to the randomness of the prediction process.

[0102] Therefore, the large-scale model for correcting thought chains needs to obtain a reference thought chain example of the first voice interaction command. Based on the reference thought chain example, the output format of the initial thought chain is adjusted to generate the target thought chain. This allows the content of the initial thought chain to be more standardized and structured without changing the original reasoning structure, thus obtaining the target thought chain.

[0103] The reference thought chain example refers to a verified COT example, which typically includes a complete, standardized, and structured description of the reasoning steps. In this application, those skilled in the art can select a batch of typical voice interaction commands based on the voice interaction scenario, write a correct and complete description of the reasoning process for each typical voice interaction command, and use this description of the reasoning process as a reference thought chain example after verification, thereby ensuring the high reliability of the reference thought chain example.

[0104] In case T0, the self-verification result of the initial thought chain is that the predicted label corresponding to the continuous sampling operation is different from the real label corresponding to the first voice interaction command. In other words, the reasoning path of the initial thought chain has a deviation.

[0105] At this point, it can be assumed that the current thought chain correction model has a significant deviation in its reasoning path for this type of instruction, such as selecting the wrong API, misjudging the meaning of the slot, or providing non-compliant interface parameters. Simply correcting the output format of the initial thought chain is no longer sufficient to achieve the desired reasoning effect.

[0106] Therefore, the large-scale model for correcting thought chains needs to determine the vehicle knowledge associated with the first voice interaction command, and based on the vehicle knowledge, perform contradiction detection on the first voice interaction command and the initial thought chain to obtain self-correction information. Based on the self-correction information, the target thought chain is regenerated.

[0107] Among them, contradiction detection refers to the process of comparing and verifying the initial thought chain generated by the thought chain correction model with voice interaction commands and vehicle knowledge. For example, it checks whether the reasoning results of the initial thought chain have logical deviations, prediction errors, or inconsistencies with vehicle knowledge, in order to determine whether the initial thought chain needs self-reflection and correction.

[0108] Self-correcting information refers to the prompts generated by the thought chain correction model based on voice interaction commands, the initial thought chain, and vehicle knowledge after a problem is detected through contradiction detection. This self-correcting information guides the thought chain correction model to generate a more reasonable SR-COT.

[0109] Optionally, the process of contradiction detection between the first voice interaction command and the initial thought chain includes at least the following: when the result generated by the initial thought chain contains an error, the erroneous result is fed back to the thought chain correction model as the first round of information; the correct result is provided to the thought chain correction model through a preset correction prompt word template; and, combined with the user command and vehicle knowledge, the model is asked to analyze the cause of the error and output a detailed thought process and the corrected result. The thought chain correction model can analyze the cause of the first-round error based on the correction prompt word template, generate self-correction information, and, based on the self-correction information, regenerate a target thought chain that conforms to the user's intent and vehicle knowledge.

[0110] As an example, the preset correction prompt template is as follows: "Your output is incorrect. The correct result is below. Please analyze the cause based on the correct result, user instructions, and vehicle knowledge, and output your thought process and final result:" [Correct Result] Selected API: {"API": "Seat_Gear_Set", "ARGUMENTS": {"device": "seat cushion", "direction": "down", "set_type": "lower", "seat_position": "third row right side"}} Based on the current correction prompt, the model will generate self-correction information, and then generate the final correct SR-COT based on the self-correction information. The generated results are as follows: [First attempt (incorrect)] Logical thinking: 1. Understanding user instructions: The user wants to increase the energy recovery level.

[0111] 2. Refer to Vehicle Knowledge: The Vehicle Knowledge does not provide detailed information on adjusting the energy recovery level, but the provided API does have settings related to energy recovery.

[0112] 3. Choose the appropriate API: The EnergyRecy_Gear_Set API allows for relative adjustment of the energy recovery level, which meets the user's needs, as the user only requires "high" energy recovery without a specific value.

[0113] 4. Fill in the API parameters: - function: 'Energy Recovery Level', because this is a function that users want to adjust.

[0114] - set_type: 'increase' indicates that the energy recovery level should be increased.

[0115] Is there a suitable API? Yes Selected API: {"API": "EnergyRecy_Gear_Set", "ARGUMENTS": {"function": "Energy Recycling Level", "set_type": "Increase"}} [Self-correction] The user command "High Energy Recovery" more accurately corresponds to setting the energy recovery level to "Large" rather than simply increasing it. Since the user hasn't provided more detailed context indicating their desired relative adjustment, setting a fixed "High Energy Recovery" level would be more appropriate.

[0116] [Second attempt (correct)] Logical thinking: 1. Understanding user instructions: The user wants to set the energy recovery level to high.

[0117] 2. Refer to Vehicle Knowledge: The Vehicle Knowledge does not provide detailed information on adjusting the energy recovery level, but the provided API does have settings related to energy recovery.

[0118] 3. Choose the appropriate API: The EnergyRecy_Max_Min_Set API allows setting the energy recovery level to fixed values ​​such as "high", "medium", and "low", which meets the user's needs because the user wants the energy recovery level to be "high".

[0119] 4. Fill in the API parameters: - function: 'Energy Recovery Level', because this is a function that users want to adjust.

[0120] - set_value: 'large', because the user wants to increase the energy recovery level, and "large" best matches the user's intent among the API's candidate values.

[0121] Is there a suitable API? Yes Selected API: {"API": "EnergyRecy_Max_Min_Set", "ARGUMENTS": {"function":"Energy Recycling Level", "set_value": "Large"}}.

[0122] In the example above, the initial thought chain was flawed on the first attempt. The thought chain correction model selected the EnergyRecy_Gear_Set API for relative adjustment, which did not fully match the user's intent. Therefore, based on a self-reflection mechanism, the thought chain correction model combined voice interaction commands and vehicle knowledge to generate self-correcting information, identifying that a fixed value setting of "high" energy recovery level should be used. Subsequently, in the second attempt, the model selected the EnergyRecy_Max_Min_Set API and set the energy recovery level to "large," thereby generating the final SR-COT that conformed to the user's intent.

[0123] In this application, the self-reflection of the initial thought chain can adopt a multi-round dialogic error correction mechanism: the thought chain correction model gradually converges its reasoning direction through continuous prompts and feedback, avoiding deviations or insufficient understanding caused by a one-time output.

[0124] The first round of dialogue focuses on quickly diagnosing and coarsely correcting the initial thought process, identifying and listing the main errors (such as selecting the wrong API, missing required slots, or non-standard parameter formats).

[0125] The second round of dialogue, based on the results of the first round of diagnosis, corrects the output format of the initial thought chain by combining reference thought chain examples, or, based on the vehicle knowledge associated with the first voice interaction command, performs contradiction detection on the first voice interaction command and the initial thought chain to obtain self-correction information, so as to facilitate re-reasoning.

[0126] As an example, if the target thought chain accurately reflects the user's intent and the API and its interface parameters are correctly selected, but the output format or reasoning structure of the target thought chain still does not meet expectations, a multi-turn dialogue can be initiated to correct the output format or reasoning structure of the target thought chain.

[0127] First round of output (initial model reasoning) In the first round, the thought chain correction model directly generates the reasoning process and API call results based on user instructions. The output at this stage is usually quite accurate in semantic understanding, but the reasoning structure and format may still be non-standard, for example: "Logic of thought: The user needs to adjust the driver's seat leg rest and lower it, but no specific value is given. Therefore, theoretically, the Seat_Value_Set API should be used, but due to the lack of value, the Seat_Gear_Set API can only be selected."

[0128] Is there a suitable API? Yes Selected API: {"API": "Seat_Gear_Set", "ARGUMENTS": {"device": "Leg rest", "direction": "Down", "set_type": "Lower", "seat_position": "Driver's seat"}} Although the API was selected correctly in the first round of output, the reasoning level was insufficient and the format did not meet the requirements, so it needs to be corrected in the next round.

[0129] Second round of input (formatting guidelines) The first round's output format or inference structure does not conform to the preset template, and the output format needs to be corrected. Please correct your results based on the error analysis above. Additionally, your output format should be similar to the format shown in the [Output Example] below. Output Example Example 1 Logical thinking: 1. Understanding user instructions: The user wants the lumbar support to be adjusted upwards.

[0130] 2. Refer to vehicle information: According to vehicle information, the lumbar support is adjustable up and down. This is available in the provided API.

[0131] 3. Choose the appropriate API: The Seat_Gear_Set API allows for relative adjustment of the device, which meets the user's needs because the user does not specify a specific value, but only requests "a little increase".

[0132] 4. Fill in the API parameters: - device: 'lumbar support', because the user specified that the lumbar support should be adjusted.

[0133] - direction: 'up', the user wants the lumbar support to move upwards.

[0134] - set_type: 'raise' means that the position of the lumbar support should be raised.

[0135] - seat_position: Since the user did not specify a particular seat, this parameter can be omitted.

[0136] Is there a suitable API? Yes Selected API: {"API": "Seat_Gear_Set", "ARGUMENTS": {"device": "lumbar support", "direction": "up", "set_type": "adjust up"}} Example 2 Logical thinking: 1. Understanding User Commands User Command: "Adjust lumbar support in the car" - The user wants to adjust the "lumbar support".

[0137] - The user wants the lumbar support to be "adjusted up".

[0138] 2. Refer to vehicle knowledge: According to the vehicle knowledge, the lumbar support can be adjusted forward and backward, as well as up and down. The user command mentioned "adjust upward," so the lumbar support needs to be adjusted upward.

[0139] 3. Choose the appropriate API According to the list of available APIs: - Seat_Max_Min_Set: Used to adjust the device's maximum and minimum values ​​(highest, lowest, first, last), which does not meet user needs.

[0140] - Seat_Gear_Set: Used to adjust the relative position of the device (higher, lower, forward, backward) to meet user needs.

[0141] - Seat_Value_Set: This is used to adjust the specific value of the device. The user command did not mention a specific value, so it does not meet the requirements.

[0142] 4. Populate API parameters Select the Seat_Gear_Set API: - device: lumbar support, because the user specified that the lumbar support should be adjusted.

[0143] - direction: up, the user wants the lumbar support to move upwards.

[0144] - set_type: Increase, indicating that the position of the lumbar support should be adjusted to be higher.

[0145] - seat_position: All (Since the user mentioned "inside the car," it means the user wants to adjust all the lumbar support positions inside the car, so "All" is selected.) Is there a suitable API? Yes Selected API: {"API": "Seat_Gear_Set", "ARGUMENTS": {"device": "lumbar support", "direction": "up", "set_type": "adjust up", "seat_position": "all"}}.

[0146] Second round output (formatted model results) Upon receiving a correction request, the model will reorganize the inference chain based on the example, generating a properly formatted and logically complete output. For example: "Logic of thought:" 1. Understanding user instructions: The user wants the driver's leg rest position to be lowered.

[0147] 2. Based on vehicle knowledge: According to vehicle knowledge, a leg rest can be understood as a seat cushion, and a seat cushion is adjustable in height. The user did not mention a specific value, but only wanted it adjusted lower.

[0148] 3. Choose the appropriate API: The Seat_Gear_Set API allows for relative adjustment of the device, which meets the user's needs because the user did not specify a specific value, but only requested to "lower it a little bit".

[0149] 4. Fill in the API parameters: - device: 'Leg rest', because the user specified that the leg rest should be adjusted.

[0150] - direction: 'down', the user wants the leg rest to move downwards.

[0151] - set_type: 'lower' means that the position of the leg rest should be lowered.

[0152] - seat_position: 'driver', because the user specified the seat position to be adjusted to be the driver's seat.

[0153] Is there a suitable API? Yes Selected API: {"API": "Seat_Gear_Set", "ARGUMENTS": {"device": "Leg rest", "direction": "Down", "set_type": "Lower", "seat_position": "Driver's seat"}}.

[0154] The final output not only maintains semantic correctness, but also fully complies with the requirements of standardized structured templates.

[0155] If the large-scale thinking chain correction model still cannot generate a target thinking chain that is consistent with the real label and meets the output format requirements after introducing human feedback and undergoing multiple rounds of self-correction, it indicates that this type of instruction is beyond the understanding range of the voice interaction model and the large-scale thinking chain correction model, and requires human intervention or should be discarded.

[0156] At the same time, record and report the complete metadata of the instruction, such as the original instruction transcription, initial thought chain, all sampling results, human feedback, correction prompts, model output and scores for each round, etc., for post-analysis.

[0157] Step 204: Combine the first voice interaction command and the initial thought chain into the first training sample.

[0158] Since the initial thought chain may contain reasoning biases, if the initial thought chain and the target thought chain are directly combined into the same training sample, the voice interaction model may learn both incorrect and correct reasoning during the learning process. This may lead to problems such as gradient conflict and confusion of reasoning patterns, thereby destroying the training effect.

[0159] Therefore, this application combines and encapsulates the first voice interaction command obtained and the initial thought chain generated in step 202 according to a predefined data format to construct the first training sample for supervised fine-tuning.

[0160] The first training sample can retain the initial thought chain for comparative training of the voice interaction model, and can also prevent the voice interaction model from learning incorrect reasoning patterns from incorrect reasoning.

[0161] Step 205: Combine the first voice interaction command, self-correction information, and target thought chain into a second training sample.

[0162] The target thought chain is accurate data generated through self-reflection (SR-COT). This application combines and encapsulates the first voice interaction command obtained and the target thought chain obtained after self-reflection in step 203 according to a predefined data format to construct a second training sample for supervised fine-tuning. This ensures that the voice interaction model updates model parameters only based on the target thought chain, avoids the impact of erroneous reasoning on the learning of the voice interaction model, and effectively improves the reasoning accuracy and self-correction ability of the voice interaction model.

[0163] The second training sample uses a target thought chain generated through self-reflection (SR-COT). The target thought chain has been verified as an accurate reasoning path and does not contain any errors or biases that may exist in the initial thought chain.

[0164] The preset data format standardizes the field composition of training samples, ensuring structural consistency across samples from different sources. For example, the first voice interaction command is written into field A, the initial thought chain or target thought chain is written into field B, and optional fields C such as real labels, error type descriptions, and NLU results are added, thus ensuring structural consistency across training data from different sources.

[0165] Step 206: Optimize the pre-configured voice interaction model based on the first training sample and the second training sample.

[0166] In this application, the thought chain correction model acts as the teacher model, while the voice interaction model to be optimized acts as the student model. Through supervised fine-tuning, the student model is made capable of self-reflection, thereby improving the prediction accuracy and reliability of the student model.

[0167] Furthermore, in the process of training the voice interaction model using the first training sample, this application does not calculate the loss value of the first training sample. In the process of training the voice interaction model using the second training sample, the loss value of the second training sample is calculated, and the model parameters of the voice interaction model are updated based on the loss value until the output result of the voice interaction model meets the preset conditions.

[0168] Here, the loss value refers to the metric that measures the difference between the model's predicted result and the actual result. In this application, the difference between the predicted label output by the voice interaction model and the actual label of the first voice interaction command is considered. The smaller the loss value, the more accurate the prediction of the voice interaction model; the larger the loss value, the worse the prediction of the voice interaction model.

[0169] Model parameters refer to the internal weights that make up a voice interaction model, determining how the model generates output from input. For example, each layer of the neural network in a voice interaction model affects the processing performance of voice commands, and updating the model parameters of each layer is actually to enable the voice interaction model to learn to generate more accurate output results under similar commands.

[0170] Preset conditions refer to the termination or judgment criteria of the training process, used to determine whether the model has achieved the expected results. For example, preset conditions may include the accuracy of the voice interaction model's output reaching a certain threshold, the proportion of the voice interaction model's output that matches the user's intent and vehicle knowledge on the validation dataset meeting requirements, or the voice interaction model's output conforming to the set business logic and having an error rate below the allowable range.

[0171] Reference Figure 4 , Figure 4 This is a schematic diagram illustrating the voice interaction model training process in an embodiment of this application.

[0172] As mentioned earlier, if the training sample containing "first attempt (error) + self-correction + second or nth attempt" is directly used to train the voice interaction model during the construction of training data, the voice interaction model will not know when the correct result will appear, causing the voice interaction model to continuously generate new trial steps, forming an output pattern similar to infinite self-iteration, resulting in unstable reasoning behavior.

[0173] Furthermore, because the training sample structure is fixed, the model assumes that "the second attempt is the final correct answer" regardless of whether the second attempt is correct. If the result of the second attempt is incorrect, the voice interaction model will learn an incorrect reasoning pattern instead of a reasonable reasoning rule.

[0174] Furthermore, the modified SR-COT (Target Thinking Chain) usually has more tokens than the first round of thinking chain. If both rounds of content are included in the loss calculation, the model will tend to retain the wrong reasoning path under gradient averaging, that is, actively generate the wrong first round of thinking chain so that it can output the correct result after entering the second round. This forms a reverse learning mode of "intentionally making a mistake first and then correcting it", which seriously affects the training stability.

[0175] To avoid the aforementioned problems, this application splits the SR-COT training process into two rounds. The first round generates the initial thought chain, and its output is not used in loss calculation. The second round generates the target thought chain after incorporating self-correction information, and its output is used only as a supervision signal for training. By designing a two-round execution structure, errors in the initial thought chain can be specifically corrected, significantly improving the model's inference accuracy.

[0176] The training process employs a Mask strategy to mask historical error information. In the two rounds of SR-COT training, the "error or potentially error COT (Initial Mind Chain)" generated in the first round is masked and does not participate in the loss calculation. This avoids the model learning incorrect inferences and solves the gradient conflict and training collapse problems caused by the simultaneous participation of error COT and correct COT in training in the traditional way.

[0177] Unlike traditional methods that involve correcting the retrieval path and performing multiple rounds of iterative reasoning after an error is discovered, this application uses only the target thought chain generated in the second round as the sole effective training supervision signal, ensuring that the model is not contaminated by erroneous reasoning in the first round. Simultaneously, without sacrificing the original reasoning and generalization abilities, it directly corrects the model's reasoning path and thinking methods, thereby significantly improving the accuracy of model reasoning without relying on repeated retrievals and external knowledge supplementation.

[0178] Through two rounds of training, the voice interaction model simulates real online reasoning behavior in the first round. Whether its output is correct or incorrect does not affect the training. In the second round, it explicitly judges whether the result of the first round meets the user's intent based on self-correction information. If not, it must actively generate the correct SR-COT and use this as the sole basis for training. Ultimately, the voice interaction model can learn a self-reflection mechanism to achieve automatic error correction without getting bogged down in endless trials, thus saving training costs.

[0179] In this embodiment, a first voice interaction command is first received, an initial thought chain corresponding to the first voice interaction command is constructed, the initial thought chain is self-verified, and a corresponding target thought chain is generated based on the self-verification result. Then, the first voice interaction command and the initial thought chain are combined into a first training sample, and the first voice interaction command, self-correction information, and target thought chain are combined into a second training sample. Finally, the pre-configured voice interaction model is optimized based on the first and second training samples. By constructing SR-COT, the reasoning ability of the voice interaction model is enhanced, enabling the voice interaction model to mimic the reasoning logic of SR-COT, thereby thinking deeply and generating more accurate reasoning results during actual voice interaction.

[0180] Corresponding to the aforementioned application function implementation method embodiments, this application also provides a voice interaction method, vehicle, and corresponding embodiments based on a voice interaction model.

[0181] Figure 5 This is a schematic diagram illustrating a voice interaction method based on a voice interaction model, as shown in an embodiment of this application.

[0182] See Figure 5 A voice interaction method based on a voice interaction model, comprising: Step 501: Real-time acquisition of the second voice interaction command; Step 502: Input the second voice interaction command into the optimized voice interaction model to generate the corresponding vehicle control task.

[0183] As an example of an embodiment of this application, the process of online voice interaction based on complex COT includes at least the following steps: real-time acquisition of second voice interaction commands, inputting the second voice interaction commands into the optimized voice interaction model, and generating corresponding vehicle control tasks.

[0184] The core of online voice interaction lies in reasoning about the second voice interaction commands initiated by the user in real time in the smart cockpit, and mapping the reasoning results into executable vehicle control tasks, thereby controlling seat components, vehicle air conditioning components, window components, etc.

[0185] In this application, the second voice interaction command refers to a voice interaction command initiated by the user in real time within the smart cockpit, which is used to perform reasoning and execute vehicle control tasks in the optimized voice interaction model.

[0186] It's worth noting that since the first voice interaction command is a historically collected and successfully responded-to command, it carries the correct label, i.e., the true label. The second voice interaction command, however, is a real-time collected command and is not pre-labeled; it carries the label predicted by the voice interaction model.

[0187] refer to Figure 6 , Figure 6 This is a schematic diagram illustrating the online voice interaction based on complex COT as shown in the embodiments of this application. The online processing flow of the second voice interaction command is as follows: During the actual inference phase, upon receiving the second voice interaction command "Turn off the in-car sound," relevant slots are extracted from the second voice interaction command based on the API description and interface parameters retrieved by RAG. The scope of effect is determined using vehicle knowledge, and the slots are standardized to "in-car = all." Combining vehicle knowledge and the API description, the most semantically matching APIs are selected from the candidate API set: "Headrest_Voice_Model_Close" and "Media_Voice_Open." Based on the interface parameter definitions of these APIs, the corresponding interface parameters are extracted from the command: {"target_function":"sound","seat_position":"all"}.

[0188] After obtaining the structured API call parameters, map them back to the original...<API,Arguments> Format. For APIs that are split during the training phase to improve the model's generalization ability, their mapping relationship with the original API is pre-stored in the configuration file, which can automatically complete the mapping and retain the corresponding parameters. The final output is still the original.<API,Arguments> The format ensures consistency between the API splitting during the training phase and the interface format during the actual deployment phase, keeping the entire processing flow seamless for downstream R&D modules and actual users.

[0189] The two most similar APIs and their interface parameters are combined with relevant vehicle knowledge to form the COT cue words required for the current inference. The inference is then completed based on the COT cue words, and the vehicle control task is output.

[0190] However, complex COT inference typically increases inference latency significantly. This application advances the natural language understanding results before COT inference, enabling the voice interaction model to have explicit API selection information before generating the COT. For example, by moving the "Selected API" field before the "Thinking Logic" field, the model can stop generating thought chain content after outputting the API selection during inference, thereby effectively reducing overall inference latency.

[0191] Regarding the methods in the above embodiments, the specific manner in which each step is performed has been described in detail in the embodiments of the method, and will not be elaborated further here.

[0192] Figure 7 This is a schematic diagram of the vehicle structure shown in the embodiments of this application.

[0193] See Figure 7 The vehicle 700 includes a memory 710 and a processor 720.

[0194] The processor 720 can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor. Memory 710 may include various types of storage units, such as system memory, read-only memory (ROM), and permanent storage devices. ROM may store static data or instructions required by the processor 720 or other modules of the computer. Permanent storage devices may be read-write storage devices. Permanent storage devices may be non-volatile storage devices that retain stored instructions and data even when the computer is powered off. In some embodiments, permanent storage devices use mass storage devices (e.g., magnetic or optical disks, flash memory) as permanent storage devices. In other embodiments, permanent storage devices may be removable storage devices (e.g., floppy disks, optical drives). System memory may be a read-write storage device or a volatile read-write storage device, such as dynamic random access memory. System memory may store some or all of the instructions and data required by the processor during operation. Furthermore, memory 710 may include any combination of computer-readable storage media, including various types of semiconductor memory chips (e.g., DRAM, SRAM, SDRAM, flash memory, programmable read-only memory), and disks and / or optical disks may also be used. In some embodiments, memory 710 may include a removable storage device that is readable and / or writable, such as a laser disc (CD), a read-only digital versatile optical disc (e.g., DVD-ROM, dual-layer DVD-ROM), a read-only Blu-ray disc, an ultra-high density optical disc, a flash memory card (e.g., SD card, mini SD card, Micro-SD card, etc.), a magnetic floppy disk, etc. Computer-readable storage media do not contain carrier waves or transient electronic signals transmitted wirelessly or via wired connections.

[0195] The memory 710 stores executable code, which, when processed by the processor 720, can cause the processor 720 to execute part or all of the methods described above.

[0196] Furthermore, the method according to this application can also be implemented as a computer program or computer program product, which includes computer program code instructions for performing some or all of the steps in the method described above.

[0197] Alternatively, this application may be implemented as a computer-readable storage medium (or a non-transitory machine-readable storage medium or a machine-readable storage medium) storing executable code (or computer program or computer instruction code) that, when executed by a processor of an electronic device (or server, etc.), causes the processor to perform part or all of the steps of the methods described above according to this application.

[0198] This application also provides a computer program product, which includes computer instructions that, when executed by a processor, implement the method described above.

[0199] The various embodiments of this application have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or improvement of the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. An optimization method for a voice interaction model, characterized in that, include: Receive the first voice interaction command; Construct the initial thought chain corresponding to the first voice interaction command; The initial thought chain is self-verified, and the corresponding target thought chain is generated based on the self-verification result; The self-check is used to identify whether there are errors in the initial thought chain; The pre-configured voice interaction model is optimized based on the target thought chain.

2. The method according to claim 1, characterized in that, The step of self-verifying the initial thought chain and generating a corresponding target thought chain based on the self-verification result includes: Extract the real labels from the first voice interaction command; The initial thought chain is sampled multiple times to obtain predicted labels; The self-verification result of the initial thought chain is determined based on the comparison relationship between the real label and the predicted label. Based on the self-verification results, the initial thought chain is subjected to hierarchical reasoning to generate the target thought chain.

3. The method according to claim 2, characterized in that, The step of performing hierarchical reasoning on the initial thought chain according to the self-verification result to generate the target thought chain includes: When there is at least one sampling operation whose predicted label is the same as the real label corresponding to the first voice interaction instruction, obtain a reference thought chain example of the first voice interaction instruction. Based on the reference thought chain example, the output format of the initial thought chain is adjusted to generate the target thought chain.

4. The method according to claim 2, characterized in that, The step of performing hierarchical reasoning on the initial thought chain according to the self-verification result to generate the target thought chain includes: When the predicted label corresponding to a continuous sampling operation is different from the real label corresponding to the first voice interaction command, the vehicle knowledge associated with the first voice interaction command is determined. Based on the vehicle knowledge, contradiction detection is performed on the first voice interaction command and the initial thought chain to obtain self-correction information; Based on the self-correcting information, the target thought chain is regenerated.

5. The method according to claim 4, characterized in that, The optimization of the pre-configured voice interaction model based on the target thought chain includes: The first voice interaction command and the initial thought chain are combined into a first training sample; The first voice interaction command, the self-correction information, and the target thought chain are combined into a second training sample. The pre-configured voice interaction model is optimized based on the first training sample and the second training sample.

6. The method according to claim 5, characterized in that, The step of optimizing the pre-configured voice interaction model based on the first training sample and the second training sample includes: During the process of training the voice interaction model using the first training sample, the loss value of the first training sample is not calculated. During the process of training the voice interaction model using the second training sample, the loss value of the second training sample is calculated, and the model parameters of the voice interaction model are updated based on the loss value until the output result of the voice interaction model meets the preset conditions.

7. The method according to claim 1, characterized in that, The construction of the initial thought chain corresponding to the first voice interaction command includes: Parse the first voice interaction command to obtain the voice interaction text and the vehicle knowledge associated with the voice interaction text; Retrieve a list of application programming interfaces (APIs) corresponding to the voice interaction text; the list of APIs includes several APIs and interface parameters corresponding to each API. The vehicle knowledge, the application programming interface, and the interface parameters are filled into a preset prompt word template to obtain the initial prompt word; The initial thought chain is generated by performing thought chain reasoning on the initial prompt words.

8. A voice interaction method based on a voice interaction model, characterized in that, The voice interaction model is obtained based on the optimization method described in claims 1-7, the method comprising: Real-time acquisition of second voice interaction commands; The second voice interaction command is input into the optimized voice interaction model to generate the corresponding vehicle control task.

9. A vehicle, characterized in that, include: processor; as well as A memory having executable code stored thereon, which, when executed by the processor, causes the processor to perform the method as described in any one of claims 1-8.

10. A computer-readable storage medium having executable code stored thereon, which, when executed by a processor of an electronic device, causes the processor to perform the method as described in any one of claims 1-8.

Images