Intelligent reasoning method based on chained reasoning and self-verification mechanism of large language model

Abstract

The invention relates to the field of big language model reasoning, in particular to an intelligent reasoning method based on big language model chain reasoning and a self-verification mechanism. The scheme comprises the steps of performing semantic analysis on a text input by a user, and extracting a problem structure and a reasoning intention; the problem is disassembled into a plurality of sub-problems, and a reasoning chain is constructed for reasoning; verification is performed after reasoning is completed; the verification result is judged, if verification passes, the reasoning result is reserved, and if verification of the reasoning result fails, the reasoning path is automatically optimized. Through a chain structure and a self-verification mechanism, the inference error rate is remarkably reduced, an inference path, a verification score and a logic basis are provided, and the user trust is improved. And multi-mode and multi-field tasks, such as knowledge questions and answers, logical reasoning and task decomposition, are supported. When verification fails, a reasoning path is automatically optimized, reasoning efficiency is improved, various large language models, knowledge bases and logic rule bases can be accessed, collaborative reasoning is achieved, and the method is suitable for multi-modal and multi-field intelligent reasoning tasks.

Description

Technical Field

[0001] This invention relates to the field of large language model reasoning, specifically to an intelligent reasoning method based on chain reasoning and self-verification mechanism of large language models. Background Technology

[0002] With the development of large language models (such as GPT, LLaMA, Qwen, etc.), they have made breakthrough progress in natural language understanding, generation and preliminary reasoning capabilities. However, current LLMs still have the following key technical bottlenecks in complex reasoning tasks: (1) Uncontrollable reasoning process: Large language models adopt an end-to-end generation method, lack a structured reasoning process, and the reasoning path is untraceable; (2) Frequent factual errors: The content generated by the model may not be consistent with the facts and lacks a self-verification mechanism; (3) Poor logical consistency: The conclusions in multi-step reasoning may be contradictory and lack consistency constraints; (4) Weak interpretability: The reasoning process is a "black box" and users cannot understand the reasoning logic; (5) Weak task decomposition ability: When facing complex tasks, it is difficult to effectively decompose and collaboratively solve them; (6) Poor regional language adaptability: It lacks the ability to process dialects, regional expressions and local knowledge.

[0003] While current large language models possess certain reasoning capabilities, they still suffer from problems such as uncontrollable reasoning paths, high factual error rates, weak interpretability, instability in multi-step reasoning, and inability to adapt to dialect input.

[0004] Especially in Sichuan, Chengdu and other regions, a large number of users use dialects for voice interaction. The dialects have complex phonological systems, vocabulary and tone features, which not only pose a great challenge to ASR, but also affect the quality of subsequent inference chains.

[0005] Therefore, there is a need for a complete intelligent reasoning method that can provide structured reasoning, self-verification, and dynamic re-reasoning, as well as handle differences in speech and text from different regions. Summary of the Invention

[0006] This invention aims to improve the controllability, interpretability, and accuracy of reasoning based on large language models, while providing speech parsing and semantic standardization capabilities for Sichuan dialect, Chengdu dialect, and other regional dialects, thereby enhancing the system's universality and robustness in real-world scenarios. This invention achieves the above objectives by adopting the following technical solution: This invention provides an intelligent reasoning method based on chained reasoning and a self-verification mechanism using large language models, including:

[0007] S1. Perform semantic parsing on the user input text to extract the question structure and reasoning intent;

[0008] S2. Break the problem down into multiple sub-problems and construct a reasoning chain to perform reasoning.

[0009] S3. Verify the reasoning after it is completed;

[0010] S4. Judge the verification result. If the verification passes, retain the reasoning result. If the verification of the reasoning result fails, automatically optimize the reasoning path.

[0011] Furthermore, step S2 specifically includes:

[0012] The problem is modeled as a chain of inference, C = {S0, S1, ..., S...} n In the formula, C represents the reasoning chain, S0 represents the original problem, and S... i For the reasoning subproblem of step i, S n For the final reasoning conclusion, each sub-problem depends on the result of the previous step;

[0013] The construction strategy includes:

[0014] Based on the chain construction of prompt word templates, a structured prompt-guided model is used to generate inference chains;

[0015] Based on the decomposition of the problem graph, the knowledge graph is used to perform semantic decomposition of the problem;

[0016] Based on the recursive decomposition of logical rules, the problem is formally logically modeled and reasoned step by step according to the rules;

[0017] The chain reasoning modeling process includes:

[0018] Each reasoning step S i The generation process is as follows:

[0019] R i =M(S) i-1 H i );

[0020] In the formula, M represents the large language model, and H... i R represents the prompt word template for step i. i Representing a large language model for S i-1 The reasoning result, S i It is made by R i-1 The generated subproblems.

[0021] Furthermore, step S3 specifically includes:

[0022] A self-verification mechanism is used to evaluate the credibility of the inference results. The self-verification mechanism includes the following verification dimensions:

[0023] Logical consistency verification:

[0024] LC(R i This indicates that the logical consistency of the reasoning result is being verified.

[0025] Fact consistency verification:

[0026] FC(R i This indicates that a knowledge base or fact database is used to verify whether the reasoning result conforms to the facts;

[0027] Model consistency verification:

[0028] k represents the number of large language models, I represents the indicator function, and MC(R) i This indicates the output of reasoning using multiple large language models for the same reasoning step;

[0029] If more than half of the outputs from multiple large language models are consistent, the inference result is considered reliable.

[0030] The final verification score is:

[0031] V(R i )=w1·LC(R i )+w2·FC(R i )+w3·MC(R i );

[0032] In the formula, w1, w2, and w3 are the weights of each dimension, w1 + w2 + w3 = 1, and v(R) i ) represents the final verification score, if V(R) i If ) < 0, then the re-inference mechanism is triggered.

[0033] Furthermore, step S4 specifically includes:

[0034] If the inference result fails to be verified, the inference path will be automatically optimized, and the following strategies will be tried until the maximum number of attempts is reached or the verification passes:

[0035] Change the prompt word template;

[0036] Guide the large language model to focus on the designated reasoning points;

[0037] The results of multiple rounds of reasoning are weighted and fused;

[0038] External constraint rules are introduced, including logical rules and trial rules;

[0039] Adjust the reasoning path based on the feedback mechanism;

[0040] The re-reasoning mechanism works as follows:

[0041]

[0042] In the formula, R represents the template for the k-th prompt word.i (k) represents the reasoning result of the i-th attempt.

[0043] Furthermore, if the user inputs Sichuan dialect voice, the process before step S1 also includes:

[0044] Convert the user's Sichuan dialect voice input into the corresponding text;

[0045] The ASR module converts the user's voice input into corresponding text, assigns a confidence score to each word in the text, sets a confidence score threshold, and marks the corresponding text that is below the confidence score threshold as requiring manual review. The text is then checked through manual review.

[0046] Generate prompt words;

[0047] The prompt words are as follows:

[0048] You are a professional linguist skilled in accurately converting Sichuan dialect texts into standard Mandarin texts. Please follow these rules:

[0049] Understanding the core meaning: Accurately understanding the true meaning of Sichuan dialect words and sentences;

[0050] Grammar and vocabulary conversion: Converting dialect words into equivalent Mandarin words and adjusting word order to standard Mandarin grammar;

[0051] Preserve semantics and tone: Retain all semantics and tone of the original sentence without any additions or deletions, including interrogative, exclamatory, and emphatic tones;

[0052] Output requirements: Output only the converted Mandarin text, without adding any explanation;

[0053] Input the text and prompt words into the large language model to obtain the corresponding Mandarin text.

[0054] Furthermore, if the Sichuan dialect speech contains Chengdu local dialect, then the Chengdu local dialect speech is acoustically enhanced. Based on the weakening of Chengdu dialect vowels, the weakening of retroflexion, and neutralization, the Chengdu dialect acoustic feature layer is constructed, and the recognition is completed using an expanded Chengdu dialect dictionary. Confidence labeling and verification are performed on the recognized text.

[0055] Furthermore, Chengdu dialect is converted into Standard Mandarin using a large language model, including:

[0056] It performs semantic equivalence conversion of dialect words, word order adjustment, tone preservation, and ambiguity resolution, and outputs standard Mandarin text.

[0057] Furthermore, when processing text using a large language model, regional proper nouns are uniformly mapped to structured entities, improving the accuracy of chain reasoning and fact verification.

[0058] Furthermore, the method also includes:

[0059] S5. Through the region-adaptive TTS (Text-To-Speech) module, it determines whether the user has a Chengdu accent based on the user's voice or vocabulary characteristics. If so, it adopts the Chengdu accent Mandarin prosody template to output more natural localized speech.

[0060] Furthermore, in step S2, the user's region is determined based on the corresponding text features, and the appropriate reasoning prompt word template, regional knowledge base, and logical rules are selected to improve reasoning stability and regional adaptability.

[0061] The beneficial effects of this invention are as follows:

[0062] This invention introduces a chain-like reasoning structure and a self-verification mechanism, which can structure complex problems into multi-step reasoning chains. A large language model is used to reason and solve each step, and a self-verification mechanism is introduced at each step to judge the logical consistency and factual accuracy of the reasoning results. Through the chain structure and self-verification mechanism, the reasoning error rate is significantly reduced.

[0063] This invention supports a dynamic re-inference mechanism, which automatically optimizes the inference path when verification fails and provides interpretable output, including inference path, verification basis, logic score and other information. It is suitable for multimodal and multi-domain intelligent reasoning tasks.

[0064] This invention can access various large language models, knowledge bases, and logic rule bases to achieve collaborative reasoning. Attached Figure Description

[0065] Figure 1 This is a flowchart of an intelligent reasoning method based on chain reasoning and self-verification mechanism of a large language model provided by the present invention;

[0066] Figure 2 This is the reasoning flowchart provided by the present invention;

[0067] Figure 3 This is a block diagram of the data interaction structure between the various modules of the inference system provided by the present invention. Detailed Implementation

[0068] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0069] This invention provides an intelligent reasoning method based on chained reasoning and self-verification mechanisms using a large language model, such as... Figure 1 and Figure 2 As shown, it includes:

[0070] S1. Perform semantic parsing on the user input text to extract the question structure and reasoning intent;

[0071] S2. Break the problem down into multiple sub-problems and construct a reasoning chain to perform reasoning.

[0072] The problem is modeled as a chain of inference, C = {S0, S1, ..., S...} n In the formula, C represents the reasoning chain, S0 represents the original problem, and S... i For the reasoning subproblem of step i, S n For the final reasoning conclusion, each sub-problem depends on the result of the previous step;

[0073] The construction strategy includes:

[0074] Based on the chain construction of prompt word templates, a structured prompt-guided model is used to generate inference chains;

[0075] Based on the decomposition of the problem graph, the knowledge graph is used to perform semantic decomposition of the problem;

[0076] Based on the recursive decomposition of logical rules, the problem is formally logically modeled and reasoned step by step according to the rules;

[0077] The chain reasoning modeling process includes:

[0078] Each reasoning step S i The generation process is as follows:

[0079] R i =M(S) i-1 H i );

[0080] In the formula, M represents the large language model, and H... i R represents the prompt word template for step i. i Representing a large language model for S i-1 The reasoning result, S i It is made by R i-1 The generated subproblems.

[0081] S3. Verify the reasoning after it is completed;

[0082] A self-verification mechanism is used to evaluate the credibility of the inference results. The self-verification mechanism includes the following verification dimensions:

[0083] Logical consistency verification:

[0084] LC(R i This indicates that the logical consistency of the reasoning result is being verified.

[0085] Fact consistency verification:

[0086] fC(R i This indicates that a knowledge base or fact database is used to verify whether the reasoning result conforms to the facts;

[0087] Model consistency verification:

[0088] k represents the number of large language models, I represents the indicator function, and MC(R) i This indicates the output of reasoning using multiple large language models for the same reasoning step;

[0089] If more than half of the outputs from multiple large language models are consistent, the inference result is considered reliable.

[0090] The final verification score is:

[0091] V(R i )=w1·LC(R i )+w2·FC(R i )+w3·MC(R i );

[0092] In the formula, w1, w2, and w3 are the weights of each dimension, w1 + w2 + w3 = 1, and V(R) = 1. i ) represents the final verification score, if V(R) i If ) < 0, then the re-inference mechanism is triggered.

[0093] S4. Judge the verification result. If the verification passes, retain the reasoning result. If the verification of the reasoning result fails, automatically optimize the reasoning path.

[0094] If the inference result fails to be verified, the inference path will be automatically optimized, and the following strategies will be tried until the maximum number of attempts is reached or the verification passes:

[0095] Change the prompt word template;

[0096] Guide the large language model to focus on the designated reasoning points;

[0097] The results of multiple rounds of reasoning are weighted and fused;

[0098] External constraint rules are introduced, including logical rules and trial rules;

[0099] Adjust the reasoning path based on the feedback mechanism;

[0100] The re-reasoning mechanism works as follows:

[0101]

[0102] In the formula, R represents the template for the k-th prompt word. i(k) This represents the reasoning result of the nth attempt.

[0103] S5. Through the region-adaptive TTS module, it determines whether the user has a Chengdu accent based on the user's voice or vocabulary characteristics. If so, it adopts the Chengdu accent Mandarin prosody template to output more natural localized speech.

[0104] Specifically, if the user inputs Sichuan dialect, then the process before step S1 includes:

[0105] Convert the user's Sichuan dialect voice input into the corresponding text;

[0106] The ASR module converts the user's voice input into corresponding text, assigns a confidence score to each word in the text, sets a confidence score threshold, and marks the corresponding text that is below the confidence score threshold as requiring manual review. The text is then checked through manual review.

[0107] Generate prompt words;

[0108] The prompt words are as follows:

[0109] You are a professional linguist skilled in accurately converting Sichuan dialect texts into standard Mandarin texts. Please follow these rules:

[0110] Understanding the core meaning: Accurately understanding the true meaning of Sichuan dialect words and sentences;

[0111] Grammar and vocabulary conversion: Converting dialect words into equivalent Mandarin words and adjusting word order to standard Mandarin grammar;

[0112] Preserve semantics and tone: Retain all semantics and tone of the original sentence without any additions or deletions, including interrogative, exclamatory, and emphatic tones;

[0113] Output requirements: Output only the converted Mandarin text, without adding any explanation;

[0114] Input the text and prompt words into the large language model to obtain the corresponding Mandarin text.

[0115] This invention also provides an intelligent reasoning method based on chained reasoning and self-verification mechanisms using a large language model, such as... Figure 3 As shown, it includes an input parsing module, an inference construction module, a chained inference module, a self-verification module, a dynamic re-inference module, and a result integration and output module. The data interaction process of each module is as follows.

[0116] The user inputs text into the input parsing module, which performs semantic parsing to extract the question structure and reasoning intent, and sends it to the reasoning construction module. The reasoning construction module breaks down the question into multiple sub-questions, constructs a reasoning chain, and performs reasoning through the chain-based reasoning module. During reasoning, a large language model is used to reason sequentially, generating intermediate reasoning results. The self-verification module verifies the reasoning results from multiple dimensions to determine their credibility. If the verification fails, the dynamic re-reasoning module automatically optimizes the reasoning path. If the verification succeeds, the result integration and output module combines the reasoning results and outputs the final answer and explainable information.

[0117] The alternative solutions of the present invention will be described in detail below.

[0118] Overall alternative:

[0119] As an alternative, the present invention may also avoid linear chain reasoning and instead represent the dependencies between subproblems by constructing graph structures (such as knowledge graphs or reasoning graphs) and perform reasoning using graph traversal or graph neural networks.

[0120] Differences from the original scheme: The inference path is non-linear and supports concurrent inference; the self-verification mechanism can be based on consistency detection of the graph structure; dynamic re-inference can be achieved through graph pruning or path reselection.

[0121] Applicable scenarios: multi-hop reasoning tasks; problem solving that highly depends on contextual relationships.

[0122] As an alternative, the present invention can also use a reinforcement learning framework (such as Deep Q-Learning) to train an inference policy model, select the best inference action based on the current problem state, and optimize the inference path through a reward mechanism.

[0123] The differences from the original scheme are: the reasoning process is automatically planned by the policy model; the self-verification mechanism can be achieved by embedding a logical consistency score in the reward function; and dynamic re-reasoning is achieved by adjusting the action sequence through the policy model.

[0124] Applicable scenarios: Reasoning tasks in dynamic environments; scenarios requiring real-time feedback and path adjustment.

[0125] Some alternatives:

[0126] As a first partial alternative, the present invention can replace the "chain reasoning module" with a "hierarchical reasoning module".

[0127] Technical principle: The problem is gradually broken down from high-level concepts into low-level reasoning tasks using a hierarchical abstraction approach, with each level calling different reasoning models or rule systems.

[0128] Advantages: More suitable for structured knowledge systems; easy to develop and maintain in a modular manner; can be combined with symbolic reasoning and neural reasoning.

[0129] As a second partial alternative, the present invention can replace the "self-verification module" with a "verification mechanism based on an external knowledge base".

[0130] Technical principle: It does not rely on internal model verification, but verifies the reasoning results by accessing external knowledge bases (such as encyclopedias, databases, and logical rule bases).

[0131] Advantages: Improves the accuracy of fact consistency; reduces reliance on the model's own capabilities; suitable for tasks with strong fact-dependent requirements.

[0132] As a third partial alternative, the present invention can replace the "dynamic re-inference module" with a "user feedback-driven re-inference mechanism".

[0133] Technical principle: When verification fails, the system prompts the user to intervene or provide feedback, and adjusts the reasoning path based on the feedback information.

[0134] Advantages: Suitable for high-risk or critical tasks; improves human-machine collaborative reasoning capabilities; reduces the system's reliance on complex path optimization algorithms.

[0135] Alternative implementations of multimodal reasoning:

[0136] Alternative approach: Adopt a unified multimodal encoder-decoder structure;

[0137] Technical principle: It does not distinguish between modalities and uses a unified multimodal processing model (such as Flamingo, KOSMOS-1) for end-to-end inference, reducing the complexity brought about by the modular structure.

[0138] Advantages: The system structure is simpler; it can learn modal fusion end-to-end; it is suitable for tasks with ambiguous modal boundaries.

[0139] The following will illustrate this with specific examples.

[0140] Example 1: Logical Reasoning Task

[0141] Question: If A is B's father and B is C's father, what is the relationship between A and C?

[0142] Reasoning chain construction:

[0143] 1. S0: A is B's father;

[0144] 2. S1: B is C's father;

[0145] 3.S2: According to transitivity, A is the grandfather of C;

[0146] 4.S3: Therefore, A is C's grandfather;

[0147] Self-verification mechanism:

[0148] 1. Check logical transitivity;

[0149] 2. Matching with common sense knowledge base;

[0150] 3. Multi-model consistency verification;

[0151] 4. The logical consistency score is 1.

[0152] Example 2: Open Domain Question Answering Task

[0153] Question: Why does Earth have four seasons?

[0154] Reasoning chain construction:

[0155] 1.S0: Why does the Earth have four seasons?

[0156] 2.S1: The Earth's axis of rotation is tilted;

[0157] 3.S2: Earth revolves around the Sun;

[0158] 4.S3: Temperature changes are caused by the different angles of sunlight at different times;

[0159] 5.S4: This is why the four seasons are formed.

[0160] Self-verification mechanism:

[0161] 1. Matching with the astronomical knowledge base;

[0162] 2. Multi-model consistency verification;

[0163] 3. The logical chain is complete, and the score is full marks.

[0164] This invention also provides a regional language and localized reasoning enhancement mechanism:

[0165] To improve the reasoning accuracy for localized users, this invention adds an extension mechanism for Chengdu dialect and other regional dialects based on Sichuan dialect processing, including:

[0166] Acoustic adaptation of Chengdu dialect: Extract typical sound changes in Chengdu dialect (vowel weakening, neutralization, rising tone) and construct a regional acoustic enhancement layer to achieve high-accuracy dialect ASR.

[0167] Dialect-Standard Mandarin Semantic Conversion: Through prompt word templates and a large language model, the semantics of dialect words are restored, word order is standardized, and tone is maintained.

[0168] Regional entity standardization: Map regional proper nouns to structured entities to improve reasoning consistency.

[0169] Regional Prompt Selection and Inference Chain Optimization: The regional recognition module automatically determines the user's region and selects the corresponding inference template to improve the stability of chain reasoning.

[0170] Regional TTS: Select the Chengdu-accented Mandarin or standard Mandarin prosody template according to regional characteristics to improve the naturalness of the output.

[0171] Among them, the examples of Chengdu regional adaptation are as follows:

[0172] Example: Voice input inference in Chengdu dialect;

[0173] User voice: "Yao bu yao de qu Tai Gu Li zheng qi zou ma?"

[0174] ASR recognition: Use the Chengdu dialect acoustic enhancement layer to recognize the dialect phonetic system;

[0175] Dialect-Mandarin conversion:

[0176] → "Yao bu yao qu Tai Gu Li yi qi zou ma?"

[0177] Regional entity standardization:

[0178] "Tai Gu Li" → Chengdu_TaikooLi;

[0179] (4) Chain reasoning: Perform task reasoning based on the standardized text; <​​​​​​​​​​​​​​​

[0185] Dialectal intonation consistency correction: Many modal particles in Sichuan dialect can alter sentence meaning, such as "ma," "sai," "oh," and "ha." The system uses modal particle vectors to correct the semantic parsing results, ensuring that the inference chain accurately reflects the original speech meaning. Figure 1 To.

[0186] Chengdu-themed reasoning template-driven chained reasoning: The system pre-sets reasoning templates for local Chengdu life scenarios, such as "asking for directions on Chunxi Road," "ordering food at a hot pot restaurant," and "traveling during a rainstorm." When location-related input is detected, the corresponding template is automatically applied, achieving more stable and efficient chained reasoning.

[0187] Collaborative error correction for Sichuan dialect vocabulary: In the speech recognition stage, the recognition results are jointly reordered by acoustic similarity and language model probability, which can effectively improve the recognition accuracy of Sichuan dialect vocabulary.

[0188] Chengdu dialect style transfer speech output: By introducing Chengdu dialect acoustic style vectors into TTS output, natural speech with Chengdu flavor can be generated, such as fast speech rate, rich interjections, and soft tone, which is suitable for localized intelligent assistant scenarios.

[0189] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. An intelligent reasoning method based on chained reasoning and self-verification mechanism using a large language model, characterized in that: include: S1. Perform semantic parsing on the user input text to extract the question structure and reasoning intent; S2. Break the problem down into multiple sub-problems and construct a reasoning chain to perform reasoning. S3. Verify the reasoning after it is completed; S4. Judge the verification result. If the verification passes, retain the reasoning result. If the verification of the reasoning result fails, automatically optimize the reasoning path.

2. The intelligent reasoning method based on chained reasoning and self-verification mechanism of a large language model according to claim 1, characterized in that, Step S2 specifically includes: The problem is modeled as a chain of inference, C = {S0, S1, ..., S...} n In the formula, C represents the reasoning chain, S0 represents the original problem, and S... i For the reasoning subproblem of step i, S n For the final reasoning conclusion, each sub-problem depends on the result of the previous step; The construction strategy includes: Based on the chain construction of prompt word templates, a structured prompt-guided model is used to generate inference chains; Based on the decomposition of the problem graph, the knowledge graph is used to perform semantic decomposition of the problem; Execution based on a regional knowledge base takes the results of semantic decomposition as input instructions and executes them through the regional knowledge base. Based on the recursive decomposition of logical rules, the problem is formally logically modeled and reasoned step by step according to the rules; The chain reasoning modeling process includes: Each reasoning step S i The generation process is as follows: R i =M(S i-1 ,H i ); In the formula, M represents the large language model, and H... i R represents the prompt word template for step i. i Representing a large language model for S i-1 The reasoning result, S i It is made by R i-1 The generated subproblems.

3. The intelligent reasoning method based on chained reasoning and self-verification mechanism of a large language model according to claim 2, characterized in that, Step S3 specifically includes: A self-verification mechanism is used to evaluate the credibility of the inference results. The self-verification mechanism includes the following verification dimensions: Logical consistency verification: LC(R i This indicates that the logical consistency of the reasoning result is being verified. Fact consistency verification: Fc(R i This indicates that a knowledge base or fact database is used to verify whether the reasoning result conforms to the facts; Model consistency verification: k represents the number of large language models, I represents the indicator function, and MC(R) i This indicates the output of reasoning using multiple large language models for the same reasoning step; If more than half of the outputs from multiple large language models are consistent, the inference result is considered reliable. The final verification score is: V(R i )=w1·LC(R i )+w2·FC(R i )+w3·MC(R i ); In the formula, w1, w2, and w3 are the weights of each dimension, w1 + w2 + w3 = 1, and V(R) = 1. i ) represents the final verification score, if V(R) i If ) < 0, then the re-inference mechanism is triggered.

4. The intelligent reasoning method based on chained reasoning and self-verification mechanism of a large language model according to claim 3, characterized in that, Step S4 specifically includes: If the inference result fails to be verified, the inference path will be automatically optimized, and the following strategies will be tried until the maximum number of attempts is reached or the verification passes: Change the prompt word template; Guide the large language model to focus on the designated reasoning points; The results of multiple rounds of reasoning are weighted and fused; External constraint rules are introduced, including logical rules and trial rules; Adjust the reasoning path based on the feedback mechanism; The re-reasoning mechanism works as follows: In the formula, R represents the template for the k-th prompt word. i(k) This represents the reasoning result of the nth attempt.

5. The intelligent reasoning method based on chained reasoning and self-verification mechanism of a large language model according to claim 1, characterized in that, If the user inputs Sichuan dialect, then the process before step S1 includes: Convert the user's Sichuan dialect speech into corresponding text; The ASR module converts the user's voice input into corresponding text, assigns a confidence score to each word in the text, sets a confidence score threshold, and marks the corresponding text that is below the confidence score threshold as requiring manual review. The text is then checked through manual review. Generate prompt words; The prompt words are as follows: You are a professional linguist skilled in accurately converting Sichuan dialect texts into standard Mandarin texts. Please follow these rules: Understanding the core meaning: Accurately understanding the true meaning of Sichuan dialect words and sentences; Grammar and vocabulary conversion: Converting dialect words into equivalent Mandarin words and adjusting word order to standard Mandarin grammar; Preserve semantics and tone: Retain all semantics and tone of the original sentence without any additions or deletions, including interrogative, exclamatory, and emphatic tones; Output requirements: Output only the converted Mandarin text, without adding any explanation; Input the text and prompt words into the large language model to obtain the corresponding Mandarin text.

6. The intelligent reasoning method based on chained reasoning and self-verification mechanism of a large language model according to claim 5, characterized in that, If the Sichuan dialect speech contains Chengdu local dialect, then the Chengdu local dialect speech is acoustically enhanced. Based on the weakening of Chengdu dialect vowels, the weakening of retroflexion, and neutralization, the Chengdu dialect acoustic feature layer is constructed, and the recognition is completed using an expanded Chengdu dialect dictionary. Confidence labeling and verification are performed on the recognized text.

7. The intelligent reasoning method based on chained reasoning and self-verification mechanism of a large language model according to claim 6, characterized in that, The Chengdu dialect was converted into Standard Mandarin using a large language model, including: It performs semantic equivalence conversion of dialect words, word order adjustment, tone preservation, and ambiguity resolution, and outputs standard Mandarin text.

8. The intelligent reasoning method based on chained reasoning and self-verification mechanism of a large language model according to claim 5, characterized in that, When processing text using a large language model, geographical proper nouns are uniformly mapped to structured entities, improving the accuracy of chain reasoning and fact verification.

9. The intelligent reasoning method based on chained reasoning and self-verification mechanism of a large language model according to claim 5, characterized in that, The method also includes: S5. Through the region-adaptive TTS module, it determines whether the user has a Chengdu accent based on the user's voice or vocabulary characteristics. If so, it adopts the Chengdu accent Mandarin prosody template to output more natural localized speech.

10. The intelligent reasoning method based on chained reasoning and self-verification mechanism of a large language model according to claim 1, characterized in that, In step S2, the user's region is determined based on the corresponding text features, and the appropriate reasoning prompt word template, regional knowledge base, and logical rules are selected to improve reasoning stability and regional adaptability.

Images