A problem solving system and method based on multi-agent collaboration

Abstract

The application discloses a kind of problem solving system and method based on multi-agent cooperation, the system is aimed at the low reliability, inefficiency, lack of error detection mechanism and other defects existing in complex scientific problem reasoning of existing large language model, proposes "decomposition-verification-conquest" architecture, includes four core agents of planner, verifier, aggregator, evaluator.Planner agent dynamically selects problem processing strategy, verifier agent executes multi-level scientific rationality verification, aggregator agent realizes scheme merging driven by dependency relationship, evaluator agent judges problem simplification degree.Experiments show that the accuracy of the system on AIME, GPQA-Diamond and other data sets is significantly improved, has high reliability, explainability and adaptability, and can be widely used in high-risk scientific reasoning scenarios.

Description

Technical Field

[0001] This invention belongs to the field of computer processing technology, specifically relating to a problem-solving system and method based on multi-agent cooperation. Background Technology

[0002] With the rapid development of Large Language Model (LLM) technology, solving complex reasoning tasks based on LLM has become an important research direction in the field of artificial intelligence. Currently, there are three main technical approaches: the first is traditional single-path reasoning methods, such as chain-of-thought reasoning, which guides the model to generate a step-by-step reasoning process to solve complex problems; the second is search-based reasoning methods, which explore multiple possible solution paths through tree search or graph search; and the third is tool-enhanced reasoning methods, which use external tools or knowledge bases to assist the reasoning process.

[0003] The aforementioned methods have demonstrated some capability in scientific reasoning tasks such as mathematics, physics, and chemistry, but they still suffer from the following key problems: First, single-path reasoning methods suffer from low reliability; once an error occurs in the reasoning path, the entire solution process fails, and it is difficult to detect and correct the error. Second, existing methods are computationally inefficient; for complex problems, they often require generating a large number of reasoning steps and intermediate results, leading to high computational costs. Third, they lack effective error detection and correction mechanisms, failing to guarantee the logical correctness and scientific rationality of the reasoning process. Fourth, existing methods struggle to handle the hierarchical decomposition of complex problems and cannot effectively manage the dependencies between subproblems. Finally, the reasoning process lacks interpretability, making it difficult to understand the model's decision-making process and verify the correctness of the results. These problems severely limit the application of LLM in high-risk, high-precision scientific reasoning tasks, necessitating a new solution system that balances reliability and efficiency. Summary of the Invention

[0004] To address the aforementioned issues, this invention proposes a problem-solving system and method based on multi-agent collaboration, enabling efficient, reliable, and interpretable solutions to complex scientific problems.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A multi-agent collaborative problem-solving system includes a planner agent, a verifier agent, an aggregator agent, an evaluator agent, and an interaction interface module. Each agent collaborates through a standardized interface. The system employs a decomposition-verification-conquest architecture to solve scientific problems, specifically including:

[0007] The planner agent receives structured problem input, maintains a dynamically expanding solution tree, selects nodes to be expanded through a heuristic node selection strategy, and adaptively decides on problem processing strategies; the processing strategies include problem equivalence transformation, problem decomposition, and direct solution;

[0008] The verifier agent performs multi-level verification on the transformation operation initiated by the planner agent. The multi-level verification includes equivalent transformation verification and problem decomposition verification. When the verification fails, feedback information containing the reason for the error is generated.

[0009] The evaluator agent performs a complexity evaluation on the leaf node problems of the solution tree to determine whether the conditions for direct solution are met.

[0010] The aggregator agent constructs a sub-problem dependency graph, uses a topological sorting algorithm to process dependencies, integrates sub-problem solutions through a recursive merging mechanism, and outputs the final answer to the original problem.

[0011] Preferably, the heuristic node selection strategy of the planner agent includes priority rules. The specific process of the priority rules is as follows: prioritize selecting nodes with a depth less than a threshold, prioritize selecting nodes with an in-degree of 0 and no dependencies, and exclude nodes that fail in all transformations. The decision logic of the processing strategy is Strategy = Planner (N, Context); where Strategy is the strategy that can be executed for the selected node; Planner (·) is the strategy that the planning agent decides to take for the selected node based on the information of the node; N is the selected node; and Context contains the solution tree state and historical processing information.

[0012] Preferably, the equivalent transformation verification dimension of the verifier agent includes the bidirectional inclusion of solutions to the original problem and the transformed problem, the domain constraint compliance of the transformation process, and the semantic fidelity of the problem essence; the problem decomposition verification dimension of the verifier agent includes the complete coverage of the solution range of the sub-problem set, the completeness of sub-problem information, the accuracy of dependency labeling, and the domain principle compliance of the decomposition strategy.

[0013] Preferably, the merging mechanism of the aggregator agent includes three stages: a dependency resolution stage to construct a subproblem dependency graph, a topological sorting stage to determine the solution order using the Kahn algorithm, and a recursive merging stage to perform solution inheritance for equivalent transformation nodes and multi-solution integration for decomposition nodes; the merging mechanism also includes convergence node optimization: when there is a unique node in the subproblems generated by decomposition that is not dependent on by other subproblems, the solution of the node is directly used as the solution of the parent problem.

[0014] Preferably, the problem-solving system based on multi-agent collaboration further includes an iterative control module, which is configured with an automatic retry mechanism: when the verifier agent fails to verify, the feedback information is passed to the planner agent as a basis for improvement, and a maximum of 3 verification attempts are performed; when the maximum number of iterations is reached or there are no available expansion nodes, all unsolved leaf nodes with an in-degree of 0 are forcibly solved.

[0015] A problem-solving method based on multi-agent cooperation, employing the aforementioned multi-agent cooperative problem-solving system, includes the following steps:

[0016] S1. Initialization: Receives the structured input of the scientific problem, creates a solution tree for the initial state, and starts the iterative solution loop; the structured input includes a set of conditions and a set of objectives.

[0017] S2, Node Expansion: The planner agent uses a heuristic node selection strategy to select nodes to be expanded, makes decisions and executes problem equivalence transformation, problem decomposition or direct solution strategies, and generates temporary child nodes;

[0018] S3, Verification Feedback: The verifier agent performs multi-level verification on temporary child nodes. If the verification passes, it is added to the solution tree. If the verification fails, feedback information is generated and retry logic is executed.

[0019] S4. Leaf Node Processing: The evaluator agent determines whether a leaf node can be solved directly, and executes a direct solution strategy for nodes that meet the conditions.

[0020] S5, Solution Merging: When the iteration terminates, the aggregator agent parses the dependencies between subproblems, recursively merges solutions after topological sorting, and outputs the final answer.

[0021] Preferably, the verification process in step S3 adopts a template-based prompting engineering and multi-temperature retry mechanism; the feedback information includes specific dimensions of verification failure and improvement suggestions.

[0022] Preferably, the iteration termination conditions in step S5 include reaching a preset maximum number of iterations and having no nodes to be expanded that meet the conditions.

[0023] By adopting the above technical solution, the present invention has the following beneficial effects:

[0024] 1. On the AIME2024 dataset, this invention improves the accuracy of the DeepSeek-V3 model from 30% to 36.67%, and the GPT-4o model from 13.33% to 16.67%; on the AIME2025 dataset, the accuracy of the DeepSeek-V3 model from 20% to 30.00%, and the GPT-4o model from 3.33% to 13.33%; on the GPQA-Diamond dataset, the accuracy of the DeepSeek-V3 model from 58.58% to 65.15%, and the GPT-4o model from 50% to 55.54%.

[0025] 2. The multi-level verification and automatic retry mechanism of this invention can reduce the risk of error propagation, achieve partial failure recovery, and has high reliability. The solution tree of this invention records the complete transformation trajectory and verification results, the reasoning process is traceable, and it has strong interpretability. Attached Figure Description

[0026] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0028] like Figure 1 As shown, a multi-agent collaborative problem-solving system includes a planner agent, a verifier agent, an aggregator agent, an evaluator agent, and an interaction interface module. Each agent collaborates through a standardized interface. The system employs a decomposition-verification-conquest architecture to solve scientific problems, specifically including:

[0029] The planner agent receives structured problem input, maintains a dynamically expanding solution tree, selects nodes to be expanded through a heuristic node selection strategy, and adaptively decides on problem processing strategies; the processing strategies include problem equivalence transformation, problem decomposition, and direct solution;

[0030] The heuristic node selection strategy of the planner agent includes priority rules. The specific process of the priority rules is as follows: prioritize selecting nodes with a depth less than a threshold, prioritize selecting nodes with an in-degree of 0 and no dependencies, and exclude nodes that fail in all transformations. The decision logic of the processing strategy is Strategy = Planner (N, Context); where Strategy is the strategy that can be executed for the selected node; Planner (·) is the strategy that the planning agent decides to take for the selected node based on the information of the node; N is the selected node; Context contains the solution tree state and historical processing information.

[0031] The verifier agent performs multi-level verification on the transformation operation initiated by the planner agent. The multi-level verification includes equivalent transformation verification and problem decomposition verification. When the verification fails, feedback information containing the reason for the error is generated.

[0032] The equivalent transformation verification dimension of the verifier agent includes the bidirectional inclusion of solutions to the original problem and the transformed problem, the domain constraint compliance of the transformation process, and the semantic fidelity of the problem essence; the problem decomposition verification dimension of the verifier agent includes the complete coverage of the solution range of the sub-problem set, the completeness of sub-problem information, the accuracy of dependency labeling, and the domain principle compliance of the decomposition strategy.

[0033] The evaluator agent performs a complexity evaluation on the leaf node problems of the solution tree to determine whether the conditions for direct solution are met.

[0034] The aggregator agent constructs a sub-problem dependency graph, uses a topological sorting algorithm to process dependencies, integrates sub-problem solutions through a recursive merging mechanism, and outputs the final answer to the original problem.

[0035] The merging mechanism of the aggregator agent includes three stages: the dependency resolution stage constructs a subproblem dependency graph; the topological sorting stage uses the Kahn algorithm to determine the solution order; and the recursive merging stage performs solution inheritance for equivalent transformation nodes and multi-solution integration for decomposition nodes. The merging mechanism also includes convergence node optimization: when there is a unique node in the subproblems generated by decomposition that is not dependent on by other subproblems, the solution of the node is directly used as the solution of the parent problem.

[0036] The problem-solving system based on multi-agent collaboration also includes an iterative control module, which is configured with an automatic retry mechanism: when the verifier agent fails to verify, the feedback information is passed to the planner agent as a basis for improvement, and a maximum of 3 verification attempts are performed; when the maximum number of iterations is reached or there are no available expansion nodes, all unsolved leaf nodes with an in-degree of 0 are forcibly solved.

[0037] A problem-solving method based on multi-agent cooperation, employing the aforementioned multi-agent cooperative problem-solving system, includes the following steps:

[0038] S1. Initialization: Receives the structured input of the scientific problem, creates a solution tree for the initial state, and starts the iterative solution loop; the structured input includes a set of conditions and a set of objectives.

[0039] S2, Node Expansion: The planner agent uses a heuristic node selection strategy to select nodes to be expanded, makes decisions and executes problem equivalence transformation, problem decomposition or direct solution strategies, and generates temporary child nodes;

[0040] S3, Verification Feedback: The verifier agent performs multi-level verification on temporary child nodes. If the verification passes, it is added to the solution tree. If the verification fails, feedback information is generated and retry logic is executed.

[0041] The verification process in step S3 employs template-based prompting engineering and a multi-temperature retry mechanism; the feedback information includes specific dimensions of verification failure and improvement suggestions.

[0042] S4. Leaf Node Processing: The evaluator agent determines whether a leaf node can be solved directly, and executes a direct solution strategy for nodes that meet the conditions.

[0043] S5, Solution Merging: When the iteration terminates, the aggregator agent resolves the dependencies between subproblems, recursively merges solutions after topological sorting, and outputs the final answer;

[0044] The iteration termination conditions in step S5 include reaching the preset maximum number of iterations and having no nodes to be expanded that meet the conditions.

[0045] Performance testing:

[0046] (I) System Deployment

[0047] This invention discloses a multi-agent collaborative problem-solving system implemented in Python. Each agent employs a modular design and interacts via a RESTful API. The LLM uses DeepSeek-V3 and GPT-4o, and the validation rule templates are customized for mathematics, physics, and chemistry respectively. Dependency graph construction utilizes the NetworkX library, and topological sorting implements the Kahn algorithm.

[0048] (II) Specific Operating Procedures

[0049] Taking the solution of an AIME math competition problem as an example, the process is as follows:

[0050] The input is a structured problem: "Given positive integers a and b that satisfy a² + b² = 25, find the maximum value of a + b", including the condition C = {a, b ∈ N*} and the objective O = {max (a + b)}.

[0051] The planner agent initializes the solution tree, selects the root node to execute the problem decomposition strategy (PD), and decomposes it into three sub-problems: "listing pairs of positive integers that satisfy a² + b² = 25", "calculating the sum of each pair", and "determining the maximum value".

[0052] The validator agent checks the completeness of the decomposition: confirming that the subproblems cover the solution objective, and the dependency relationship is marked as "subproblem 1 → subproblem 2 → subproblem 3", the verification is successful;

[0053] The evaluator agent determines that subproblem 1 can be solved directly, and outputs pairs (3,4), (4,3), and (5,0) (excluding non-positive integer pairs);

[0054] The aggregator agent calculates subproblem 2 (sum of 7, 7, 5) and subproblem 3 (maximum value 7) in topological order, and then merges them to obtain the final answer 7.

[0055] (III) Experimental Verification

[0056] The test results on the AIME2024, AIME2025, and GPQA-Diamond datasets are shown in Table 1 below:

[0057] Table 1: Test results of the multi-agent collaborative problem-solving system of the present invention on AIME2024, AIME2025 and GPQA-Diamond datasets.

[0058]

[0059] Experiments show that the present invention achieves a significant improvement in accuracy, verifying the effectiveness of the technical solution. The multi-agent collaborative problem-solving system of the present invention maintains higher accuracy, mainly due to: (1) accurate problem decomposition avoiding redundant reasoning steps; (2) early error detection by the Verifier Agent avoiding in-depth exploration of erroneous paths; and (3) optimization of the aggregation node by the Aggregator Agent reducing unnecessary merging operations. In terms of reliability, the multi-level verification mechanism significantly reduces the propagation of reasoning errors. The rigorous checks by the Aggregator Agent ensure that each transformation step is scientifically reasonable, and the error recovery mechanism enables the system to automatically recover from local failures. In terms of interpretability, the tree-structured solution process provides a clear reasoning trajectory, and each node records detailed transformation information and verification results, facilitating understanding and debugging. In terms of adaptability, the modular architecture design enables the system to easily adapt to different scientific fields and problem types, requiring only adjustments to the corresponding prompt templates and verification rules. These technical effects demonstrate that the present invention is not only theoretically innovative, but also exhibits significant technical advantages and practical value in practical applications, providing a new technical path for the automated solution of complex scientific problems.

[0060] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A problem-solving system based on multi-agent cooperation, characterized in that, The system includes a planner agent, a verifier agent, an aggregator agent, an evaluator agent, and an interaction interface module. Each agent collaborates through a standardized interface. The multi-agent collaborative problem-solving system employs a decomposition-verification-conquest architecture to solve scientific problems, specifically including: The planner agent receives structured problem input, maintains a dynamically expanding solution tree, selects nodes to be expanded through a heuristic node selection strategy, and adaptively decides on problem processing strategies; the processing strategies include problem equivalence transformation, problem decomposition, and direct solution; The heuristic node selection strategy of the planner agent includes priority rules. The specific process of the priority rules is as follows: prioritize selecting nodes with a depth less than a threshold, prioritize selecting nodes with an in-degree of 0 and no dependencies, and exclude nodes that fail in all transformations. The decision logic of the processing strategy is Strategy = Planner (N, Context); where Strategy is the strategy that can be executed for the selected node; Planner (·) is the strategy that the planning agent decides to take for the selected node based on the information of the node; N is the selected node; Context contains the solution tree state and historical processing information. The verifier agent performs multi-level verification on the transformation operation initiated by the planner agent. The multi-level verification includes equivalent transformation verification and problem decomposition verification. When the verification fails, feedback information containing the reason for the error is generated. The equivalent transformation verification dimension of the verifier agent includes the bidirectional inclusion of solutions to the original problem and the transformed problem, the domain constraint compliance of the transformation process, and the semantic fidelity of the problem essence; the problem decomposition verification dimension of the verifier agent includes the complete coverage of the solution range of the sub-problem set, the completeness of sub-problem information, the accuracy of dependency labeling, and the domain principle compliance of the decomposition strategy. The evaluator agent performs a complexity evaluation on the leaf node problems of the solution tree to determine whether the conditions for direct solution are met. The aggregator agent constructs a sub-problem dependency graph, uses a topological sorting algorithm to process dependencies, integrates sub-problem solutions through a recursive merging mechanism, and outputs the final answer to the original problem. The merging mechanism of the aggregator agent includes three stages: the dependency resolution stage constructs a subproblem dependency graph; the topological sorting stage uses the Kahn algorithm to determine the solution order; and the recursive merging stage performs solution inheritance for equivalent transformation nodes and multi-solution integration for decomposition nodes. The merging mechanism also includes convergence node optimization: when there is a unique node in the subproblems generated by decomposition that is not dependent on by other subproblems, the solution of the node is directly used as the solution of the parent problem.

2. The problem-solving system based on multi-agent cooperation as described in claim 1, characterized in that, The problem-solving system based on multi-agent collaboration also includes an iterative control module, which is configured with an automatic retry mechanism: when the verifier agent fails to verify, the feedback information is passed to the planner agent as a basis for improvement, and a maximum of 3 verification attempts are performed; when the maximum number of iterations is reached or there are no available expansion nodes, all unsolved leaf nodes with an in-degree of 0 are forcibly solved.

3. A problem-solving method based on multi-agent cooperation, characterized in that, The problem-solving system based on multi-agent cooperation as described in any one of claims 1-2 includes the following steps: S1. Initialization: Receives the structured input of the scientific problem, creates a solution tree for the initial state, and starts the iterative solution loop; the structured input includes a set of conditions and a set of objectives. S2, Node Expansion: The planner agent uses a heuristic node selection strategy to select nodes to be expanded, makes decisions and executes problem equivalence transformation, problem decomposition or direct solution strategies, and generates temporary child nodes; S3, Verification Feedback: The verifier agent performs multi-level verification on temporary child nodes. If the verification passes, it is added to the solution tree. If the verification fails, feedback information is generated and retry logic is executed. S4. Leaf Node Processing: The evaluator agent determines whether a leaf node can be solved directly, and executes a direct solution strategy for nodes that meet the conditions. S5, Solution Merging: When the iteration terminates, the aggregator agent parses the dependencies between subproblems, recursively merges solutions after topological sorting, and outputs the final answer.

4. The problem-solving method based on multi-agent cooperation as described in claim 3, characterized in that: The verification process in step S3 adopts a template-based prompting engineering and multi-temperature retry mechanism; the feedback information includes specific dimensions of verification failure and improvement suggestions.

5. The problem-solving method based on multi-agent cooperation as described in claim 3, characterized in that: The iteration termination conditions in step S5 include reaching the preset maximum number of iterations and having no nodes to be expanded that meet the conditions.

Images