Multi-thinking chain strategy method based on large language model progressive reasoning
Through the progressive reasoning framework of multi-agent collaboration, the human cognitive system is simulated, multiple propositions are generated and logical consistency checks are carried out, which solves the multi-factor decomposition and intersection problems of large language models in complex reasoning problems, and improves the reasoning ability.
Patent Information
- Application Number
- CN202510250575.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-04
- Publication Date
- 2025-08-01
AI Technical Summary
When facing complex inference problems, existing large language models have problems such as insufficient multi-factor decomposition, cross-cutting problems after decomposition of reasoning tasks, and limited improvement in reasoning capabilities.
The progressive reasoning framework of multi-agent collaboration is adopted to simulate the human cognitive system. By reviewing the reasoning module and the proposition generation module, multiple propositions are generated and logical consistency checks are performed, and the final reasoning results are generated based on known premises, breaking through the limitations of the traditional single-threaded reasoning paradigm.
It effectively reduces the interference of redundant steps in multi-step inference scenarios, ensures the diversity and comprehensiveness of inference paths, solves the problem of cross-propositions after multi-factor decomposition, and improves the reasoning ability of large language models in complex problems.
Smart Images
Figure CN120409663A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of complex problem reasoning of large language models, and specifically relates to a multi-thinking chain strategy method based on progressive reasoning of large language models. Background Art
[0002] Large language models (LLMs) have recently been shown to provide impressive performance in a variety of natural language processing tasks. For example, when the models are large enough, they exhibit reasoning capabilities. Although large language models are powerful and versatile, they often fail to solve highly complex reasoning problems. By exploring the reasoning capabilities of large language models, the needs of various practical applications can be better met, and there is broad application space for improving the reasoning capabilities of large models.
[0003] With the increasing complexity of reasoning problems and the higher requirements for reasoning accuracy, the existing large language model reasoning methods can no longer meet the high-accuracy requirements of many scenarios.
[0004] Currently, the inference of large language models only predicts the next character or word according to the given context order, rather than pausing to think, thus showing characteristics such as fast, instinctive, and emotional. This is also an important reason for their many limitations when facing complex tasks. At present, the main inference methods based on large language models are CoT, MCR, and CR. CoT involves providing some examples of "chains of thought" in the LLMs prompt, which are intermediate natural language reasoning steps. Specifically, in CoT, <input, chain of thought, output> triples are used to replace <input, output> examples. In this way, the large language model can learn the thought process of the chain of thought, so as to generate an explicit reasoning process for a given target problem. Although CoT has achieved relatively accurate reasoning ability, assuming that there can be multiple ways of thinking to deduce the final answer for each complex problem, the traditional CoT still has the problem of multi-factor decomposition of reasoning. At present, there are various methods to explore whether this idea can be used to improve the reasoning ability of large models for complex problems. MCR performs meta-reasoning through multiple reasoning chains and generates the final answer and explanation. It mainly consists of three major modules: decoding, retrieval, and meta-reasoning. The context of multiple chains and the original problem are given to the large language model to train its ability to obtain valuable information from the context of multiple chains and generate the final answer. Multi-chain-of-thought reasoning actively decomposes and reasons about the task on multiple paths, but still faces problems such as whether there is crossover after the reasoning task is decomposed and whether the reasoning ability is improved after the task decomposition crossover. CR uses 3 different large language models as the proposer, verifier, and reporter. These three models work together to simulate the human thinking process. Cumulative reasoning in CR greatly enhances the ability of the language model to handle complex tasks by decomposing each task into atomic and manageable steps. Compared with the thought tree and thought graph, cumulative reasoning is more general because it reduces the redundant steps of the thought tree and thought graph, but it still faces the dilemma that the chain of thought is limited by the linear thinking structure, and also faces the problem of whether there is crossover after the reasoning task is decomposed. In addition, although CR simulates the human thinking process based on the chain of thought, there is still room for improvement to be closer to human reasoning.
[0005] To sum up, we believe that the existing methods still need to be continuously improved and refined for problems such as multi-factor decomposition of reasoning problems, whether there is crossover after the problem reasoning task is decomposed, and whether the reasoning ability is improved after the task decomposition crossover. In addition, there is still room for improvement in the multi-agent simulation of human thinking in the CR method. Summary of the Invention
[0006] In view of the problems of the existing methods for multi-factor decomposition of reasoning problems, whether there is crossover after the problem reasoning task is decomposed, and whether the reasoning ability is improved after the task decomposition crossover, the present invention provides a multi-chain-of-thought strategy method based on progressive inference of large language models.
[0007] To achieve the above object, the present invention adopts the following technical solutions:
[0008] A multi-thinking chain strategy method based on progressive reasoning of large language models, the method comprising the following steps:
[0009] Step 1: Construct a progressive reasoning framework for multi-agent collaboration, simulate the analytical thinking mechanism in the human cognitive system, and realize the reasoning of complex problems through the collaborative work of multi-agents;
[0010] Further, the progressive reasoning framework for multi-agent collaboration in Step 1 includes a review reasoning module and a proposition generation module, and the expression is shown in Equation (1):
[0011] R = f reason (f gen (P, Q, θ), P, Q) (1)
[0012] Wherein, R is the final reasoning result, P is the major premise, Q is the input problem, θ is the generation parameter of the thinking chain, f reason and f gen are the functions of the review reasoning module and the proposition generation module respectively, and both are implemented based on the large language model;
[0013] The review reasoning module: used to review the generated propositions, and combine the reviewed and verified propositions to generate reasoning results according to the known premises;
[0014] The proposition generation module: used to generate propositions, i.e., minor premises, according to the background information provided by the problem and the known premises, i.e., the major premise.
[0015] Extract content from different perspectives in the major premise, find the key information in the dataset in multiple aspects, make the reasoning paths of the thinking chain diverse, and then generate multiple propositions to ensure the diversity and comprehensiveness of the reasoning paths; the formula is as follows:
[0016] S i = f gen (P, Q, θ i ), i = 1, 2,..., n (2)
[0017] Wherein, S i is the proposition generated by the i-th thinking chain, and θ i represents the generation parameter of the i-th thinking chain.
[0018] The review reasoning module includes the following steps:
[0019] Proposition review and validity determination, verify whether the generated proposition conflicts or collides with the known premises through logical consistency checking, and the logical consistency checking is performed by separately querying multiple large language models;
[0020] In the proposition review and validity determination, if a certain proposition conflicts with the known premises or cannot be accepted through logical reasoning, it is screened out.
[0021] Proposition content regression and integration, according to the reviewed propositions, combined with the reasoning questions and known premises, conduct detailed analysis and reasoning on the connotations of each proposition multiple times through the self-consistent thinking chain method;
[0022] Inference path explanation and result generation, use the large language model to sum up the above analysis and reasoning, and generate the inference path and result described in natural language according to formula (3):
[0023] Report(R) = Path(R) + Conclusion(R) (3)
[0024] Among them, Report(R) is the report of the result, Path(R) is the inference path, and Conclusion(R) is the summary of the inference result;
[0025] In the inference path explanation and result generation, the review and inference module combines the context, known premises and generated propositions to generate an inference process and conclusion that conform to human language habits, and reports and statistics the inference results.
[0026] Step 2: Input the problem to be processed and the known premises into the constructed progressive inference framework for multi-agent collaboration, and reason about the problem to be processed through the progressive inference framework;
[0027] Step 3: Output the inference result of the problem to be processed, and the inference result is generated by reasoning through the progressive inference framework for multi-agent collaboration;
[0028] The progressive inference framework for multi-agent collaboration is simulated and implemented by the large language model for the syllogism paradigm;
[0029] The progressive inference framework is applicable to the reasoning of complex problems and can break through the limitations of the traditional single-threaded inference paradigm.
[0030] A computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the steps of a multi-thinking chain strategy method based on progressive inference of a large language model.
[0031] An electronic device includes a processor and a memory. A computer program is stored on the memory. When the computer program is executed by the processor, it implements the steps of a multi-thinking chain strategy method based on progressive reasoning of a large language model.
[0032] Compared with the prior art, the present invention has the following advantages:
[0033] (1) The progressive reasoning framework of multi-agent collaboration proposed by the present invention can effectively reduce the interference of diverse and complex redundant reasoning steps in multi-step reasoning scenarios, such as those in the thinking tree and thinking map methods, on the reasoning results, and is helpful for the application of large language models in multi-step reasoning task scenarios;
[0034] (2) The multi-thinking chain proposition generation module proposed by the present invention, as a scheme for proposing propositions by a progressive reasoning proposer, thinks about problems from multiple perspectives, decomposes the reasoning task into multiple factors, and also solves the problem of whether there are intersections in the propositions after multi-factor decomposition in combination with the reasoning framework. BRIEF DESCRIPTION OF THE DRAWINGS
[0035] Figure 1 is a flowchart of a multi-thinking chain strategy method based on progressive reasoning of a large language model;
[0036] Figure 2 is a structural diagram of a proposition generation module of a multi-thinking chain strategy method based on progressive reasoning of a large language model;
[0037] Figure 3 is a structural diagram of a review reasoning module of a multi-thinking chain strategy method based on progressive reasoning of a large language model;
[0038] Figure 4 is a schematic diagram of an example of a reasoning result obtained by reasoning through the method in an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS
[0039] To understand the present invention in depth, we will describe it comprehensively and meticulously. However, the present invention has multiple implementation manners and is not limited to the specific examples listed herein. The presentation of these examples aims to deepen the comprehensive understanding of the disclosed content of the present invention.
[0040] A multi-thinking chain strategy method based on progressive reasoning of a large language model, the method includes the following steps, as Figure 1 shown:
[0041] Step 1: Construct a progressive reasoning framework of multi-agent collaboration, simulate the analytical thinking mechanism in the human cognitive system, and realize the reasoning of complex problems through the collaborative work of multi-agents;
[0042] Furthermore, the multi-agent collaborative progressive reasoning framework in step 1 includes a review reasoning module and a proposition generation module, and the expression is shown in formula (1):
[0043] R=f reason (f gen (P,Q,θ),P,Q) (1)
[0044] Among them, R is the final reasoning result, P is the major premise, Q is the input question, θ is the generation parameter of the thinking chain, and f reason and f gen They are the functions of the review reasoning module and the proposition generation module, both of which are implemented based on the large language model;
[0045] The review reasoning module is used to review the generated propositions and generate reasoning results based on the known premises in combination with the verified propositions, such as Figure 3 As shown;
[0046] The proposition generation module is used to generate a proposition (minor premise) based on the background information provided by the question and the known premise (major premise), such as Figure 2 shown.
[0047] Extract content from different perspectives in the major premise, find key information in the data set from multiple perspectives, diversify the reasoning paths of the thinking chain, and then generate multiple propositions to ensure the diversity and comprehensiveness of the reasoning paths; the public announcement is as follows:
[0048] S i =f gen (P,Q,θ i ),i=1,2,...,n (2)
[0049] Among them, S i is the proposition generated by the i-th thinking chain, θ i Represents the generation parameters of the i-th thinking chain.
[0050] The review reasoning module includes the following steps:
[0051] Proposition review and validity determination: verify whether the generated proposition is inconsistent or conflicting with the known premise through logical consistency checks. The logical consistency checks are performed by querying multiple large language models separately.
[0052] In the proposition review and validity judgment, if a proposition conflicts with a known premise or cannot be accepted through logical reasoning, it will be screened out.
[0053] Regression and integration of proposition content: Based on the examined propositions, combined with reasoning questions and known premises, the connotation of each proposition is repeatedly analyzed and reasoned through the self-consistent thinking chain method;
[0054] Inference path explanation and result generation: Using a large language model to sum up the above analysis and reasoning, and generating an inference path and result described in natural language according to Equation (3):
[0055] Report(R) = Path(R) + Conclusion(R) (3)
[0056] Among them, Report(R) is the report of the result, Path(R) is the inference path, and Conclusion(R) is the summary of the inference result;
[0057] In the above-mentioned inference path explanation and result generation, the review inference module combines the context, known premises and generated propositions to generate an inference process and conclusion that conform to human language habits, and reports and statistics the inference results.
[0058] Step 2: Input the problem to be processed and the known premises into the constructed progressive inference framework for multi-agent collaboration, and reason about the problem to be processed through the progressive inference framework;
[0059] Step 3: Output the inference result of the problem to be processed, and the inference result is generated by reasoning through the progressive inference framework for multi-agent collaboration;
[0060] The progressive inference framework for multi-agent collaboration is simulated and implemented through a large language model for the syllogism paradigm;
[0061] The progressive inference framework is applicable to the reasoning of complex problems and can break through the limitations of traditional single-threaded inference paradigms.
[0062] A computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it realizes the steps of a multi-thinking chain strategy method based on progressive inference of a large language model.
[0063] An electronic device, including a processor and a memory, on which a computer program is stored, and when the computer program is executed by the processor, it realizes the steps of a multi-thinking chain strategy method based on progressive inference of a large language model.
[0064] Embodiment
[0065] Step 1: Construct a progressive inference framework for multi-agent collaboration to simulate the analytical thinking mechanism in the human cognitive system;
[0066] Step 2: Input the problem to be processed and the known premises into the constructed progressive inference framework for multi-agent collaboration, and reason about the problem to be processed through the progressive inference framework;
[0067] Input the problem to be processed and the known premises, and reason about the problem through the progressive reasoning framework. For example: Input "Conclusion": "All A have C.", "Premise conditions": ["No A is B.", "All B have C."];
[0068] Combine Figure 2 As shown, generate multiple propositions through the same large language model to ensure the diversity and comprehensiveness of the reasoning path. For example, extract the subject to generate 2 propositions: "There exists an A that does not have C; No A has C", and extract the object to generate 2 propositions: "What has C is not A; There does not exist a B that does not have C".
[0069] Regression and integration of proposition content: For example, after the above 2 generated propositions pass the review and validity determination, they are regressed and integrated into 1 proposition: "There exists an A that does not have C; There does not exist a B that does not have C";
[0070] Explanation of the reasoning path and result generation: The generated explanation and result are as follows: "From the premises, we know that no A is B, all B have C, and there does not exist a B that does not have C. This means that there exists an A that does not have C, but the conclusion is that all A have C. Therefore, it cannot be determined whether this conclusion is correct or incorrect - unknow". Figure 4 This is the full process demonstration of the method's reasoning and reporting for the reasoning problem. Compared with the traditional method, the reasoning steps and proposition generation of this method decompose the reasoning task into multiple factors, and the combination of the reasoning framework also solves the problem of whether there is an intersection of propositions after the multi-factor decomposition.
[0071] Step 3: Output the reasoning result of the problem to be processed, and the reasoning result is generated through the progressive reasoning framework of multi-agent collaboration.
[0072] The content not described in detail in the specification of the present invention belongs to the prior art well-known to those skilled in the art. Although the illustrative specific embodiments of the present invention have been described above for the convenience of those skilled in the art to understand the present invention, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention defined and determined by the appended claims, these changes are obvious, and all inventions and creations using the concept of the present invention are within the scope of protection.
Claims
1. A multi-thinking chain strategy method based on progressive reasoning of large language models, characterized in that, The method includes the following steps: Step 1: Construct a progressive reasoning framework for multi-agent collaboration to simulate the analytical thinking mechanism in the human cognitive system; Step 2: Input the problem to be processed and the known premises into the constructed progressive reasoning framework for multi-agent collaboration, and reason about the problem to be processed through the progressive reasoning framework; Step 3: Output the reasoning result of the problem to be processed, where the reasoning result is generated by the progressive reasoning framework for multi-agent collaboration.
2. The multi-thinking-chain strategy method based on progressive reasoning of a large language model according to claim 1, wherein The progressive reasoning framework for multi-agent collaboration in Step 1 includes a review reasoning module and a proposition generation module, and the expression is as shown in Equation (1): R = f reason (f gen (P, Q, θ), P, Q) (1) Among them, R is the final reasoning result, P is the major premise, Q is the input question, θ is the generation parameter of the thought chain, f reason and f gen are the functions of the review reasoning module and the proposition generation module respectively; The review reasoning module: used to review the generated propositions, and combine the reviewed and verified propositions to generate a reasoning result based on the known premises; The proposition generation module: used to generate propositions (i.e., minor premises) according to the background information provided by the problem and the known premises (i.e., major premises).
3. The multi-thinking-chain strategy method based on progressive reasoning of a large language model according to claim 2, wherein, The proposition generation module includes the following steps: Extract content from different perspectives in the major premise, find the key information in the dataset in multiple aspects, make the reasoning paths of the thought chain diverse, and then generate multiple propositions, as shown below: S i = f gen (P, Q, θ i ), i = 1, 2,..., n (2) Among them, S i is the proposition generated by the i-th chain of thought, and θ i represents the generation parameter of the i-th chain of thought.
4. The multi-thinking-chain strategy method based on progressive reasoning of a large language model according to claim 2, characterized in that, The review reasoning module includes the following steps: Proposition review and validity determination, verify whether the generated proposition conflicts with the known premises through logical consistency checking, and the logical consistency checking is performed by separately asking multiple large language models; Proposition content regression and integration, according to the reviewed propositions, combine the reasoning problem and the known premises, and conduct detailed analysis and reasoning on the connotations of each proposition multiple times through the self-consistent thought chain method; Reasoning path explanation and result generation, use the large language model to sum up the above analysis and reasoning, and generate the reasoning path and result described in natural language according to Equation (3): Report(R)=Path(R)+Conclusion(R) (3) Among them, Report(R) is the report of the result, Path(R) is the reasoning path, and Conclusion(R) is the summary of the reasoning result.
5. A multi-thinking-chain strategy method based on progressive reasoning of a large language model according to claim 4, characterized in that In the proposition review and validity determination, if a certain proposition conflicts with the known premises or cannot be accepted through logical reasoning, it is screened out.
6. The multi-thinking-chain strategy method based on progressive reasoning of a large language model according to claim 4, characterized in that, In the reasoning path explanation and result generation, the review reasoning module combines the context, known premises and generated propositions to generate a reasoning process and conclusion that conform to human language habits, and reports and statistics the reasoning results.
7. A multi-thinking-chain strategy method based on progressive reasoning of a large language model according to claim 1, characterized in that, The progressive reasoning framework for multi-agent collaboration simulates and realizes the syllogism paradigm through a large language model.
8. A computer-readable storage medium, characterized in that, A computer program is stored on the storage medium, and when the computer program is executed by a processor, the steps of a multi-thought chain strategy method for progressive reasoning based on a large language model described in any one of claims 1 to 7 are implemented.
9. An electronic device, characterized in that, It includes a processor and a memory. A computer program is stored on the memory, and when the computer program is executed by the processor, the steps of a multi-thought chain strategy method for progressive reasoning based on a large language model described in any one of claims 1 to 7 are implemented.