Large model answer generation method and system based on factual reasoning

Abstract

The invention relates to a factual reasoning-based large model answer generation method and system. According to the method, after the answer draft is generated, an auditing process is carried out, firstly, a question is put forward for the answer draft, then fact checking is carried out based on the question, evidence is provided, finally, the answer draft is decided based on the evidence and the question, a deciding result is obtained, and if the deciding result comprises an accepting instruction, the current answer draft is used as a final draft; if the judgment result comprises a rejection instruction, modifying the current answer draft and re-executing the reviewing process; if the judgment result comprises a re-planning instruction, regenerating an answer draft, and performing an auditing process; and finally, optimizing the final draft, and obtaining and outputting an answer text. Compared with the prior art, the method has the advantages of remarkably improving the fact accuracy and logic preciseness of the output content and the like.

Description

Technical Field

[0001] This invention relates to the field of large model technology, and in particular to a method and system for generating large model answers based on factual reasoning. Background Technology

[0002] In recent years, artificial intelligence technologies, represented by Large Language Models (LLMs), have developed rapidly and demonstrated remarkable capabilities in fields such as natural language processing, question answering, and content generation. These models, through pre-training on massive amounts of text data, have learned rich linguistic patterns and world knowledge. However, ensuring the accuracy, reliability, and rationality of the output of large language models when processing queries requiring precise facts and rigorous logic has become a key challenge in this field. Detecting and improving the factual reasoning capabilities of these models is a core technological bottleneck driving their application in serious scenarios such as finance, law, and scientific research.

[0003] Currently, there are two main technical solutions for detecting and applying the factual reasoning capabilities of large language models. The first is the Direct Generation approach, which is the most basic application mode. The system directly inputs the user's question into the large language model, and the model, relying on its internal parameterized knowledge, generates the answer directly through a single forward computation. The entire process does not involve external information retrieval or self-correction. The second is the Retrieval-Augmented Generation (RAG) approach, an improvement on the former. Before generating the answer, this approach retrieves relevant information from an external knowledge base based on the user's question. This information is then input into the model as additional context along with the original question, guiding the model to generate a richer and more factually supported answer.

[0004] For example, the invention patent with publication number CN117786091A discloses a self-inspired intelligent question-answering method and system based on Socratic questioning. Although it automatically asks follow-up questions and answers the questions, like the search enhancement generation scheme, it still has a large number of factual errors that cannot be detected by itself due to misunderstanding of context or inherent bias, resulting in answers that are not rigorous enough.

[0005] However, both of the aforementioned existing technical solutions have significant drawbacks in practical applications, primarily stemming from their "one-time generation" paradigm. First, they generally lack a process-oriented self-examination and correction mechanism. During a single generation process, the model cannot proactively and iteratively examine and correct its intermediate thinking and final conclusions. This means that even with external knowledge references (RAG), the model may produce undetectable factual errors due to misunderstanding of context or inherent biases—a phenomenon known as "model illusion." Second, because the model's reasoning process is like a "black box," existing technologies struggle to reveal the internal reasoning chain when errors occur, making it impossible to pinpoint the specific point of error. This results in detection capabilities remaining only at the surface level of the final result, failing to deeply expose the model's specific shortcomings in multi-step reasoning or causal analysis. Finally, for complex and open-ended problems requiring comprehensive analysis or multi-step deduction, this one-time generation model struggles to construct a structured thinking framework, often leading to one-sided, logically disjointed, or overly simplified answers, failing to demonstrate rigorous and comprehensive analytical capabilities.

[0006] In summary, current technologies suffer from several drawbacks: a lack of process-based self-correction capabilities, a tendency to generate model illusions that are difficult to detect; opaque reasoning processes that make it difficult to pinpoint flaws; insufficient analytical capabilities for complex problems, difficulty in constructing structured thinking frameworks, and significant shortcomings such as one-sided answers and logical jumps. Consequently, the factual accuracy and logical rigor of the model's output are relatively low. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of the existing technology and provide a method and system for generating large model answers based on factual reasoning.

[0008] The objective of this invention can be achieved through the following technical solutions: According to one aspect of the present invention, a method for generating large-scale model answers based on factual reasoning is provided, the method comprising the following steps: S1. Perform an internal complexity analysis on the input problem and classify the problem as a simple problem or a complex problem based on the analysis results; S2. If the current problem is a simple problem, then generate a draft answer directly for the input problem and proceed to step S3; if the current problem is a complex problem, then plan, generate and synthesize a draft answer from multiple perspectives for the input problem. S3. Raise questions about the draft answer, conduct fact-checking based on the questions and provide evidence, and finally make a ruling on the draft answer based on the evidence and questions to obtain a ruling result. If the ruling result includes an acceptance instruction, the current draft answer is taken as the final draft and step S4 is executed; if the ruling result includes a rejection instruction, the current draft answer is modified and step S3 is executed again; if the ruling result includes a replanning instruction, the process jumps to step S2. S4. Optimize the final draft and output it as the answer text.

[0009] As a preferred technical solution, the specific process of performing intrinsic complexity analysis on the input problem in S1 includes: Preset inference jump threshold; Call a large language model instance, use the preset inference hop count evaluation instruction to break down the problem into inference steps, and count the number of inference hops required for the problem; If the number of reasoning hops required for a problem is less than or equal to a preset reasoning hops threshold, the problem is classified as a simple problem; otherwise, it is classified as a complex problem.

[0010] As a preferred technical solution, the specific process of multi-perspective planning, generation, and synthesis to obtain a draft answer in S2 includes: For the input problem, a structured solution outline is developed by combining domain knowledge and problem analysis dimensions; Based on the solution outline, multiple large language model instances are invoked in parallel. Each instance receives the same solution outline and is configured with different perspective instructions, thereby generating multiple drafts from multiple perspectives. Receive all generated drafts, identify them through text semantic analysis, and integrate all drafts into a consensus draft as the answer draft based on the consensus content, differences and unique information among the drafts; Among them, the multiple perspectives include a neutral perspective, a critical perspective that focuses on risk, and a quantitative perspective that focuses on data.

[0011] As a preferred technical solution, when questioning the draft answer in S3, the questions raised include factual questions, logical questions, causal questions, and adversarial questions, and the questions raised are marked with priority.

[0012] As a preferred technical solution, when S3 performs fact-checking and provides evidence based on the challenge, it specifically involves: calling the large language model itself or an external API to perform fact-checking and generating an evidence report containing the verification results; the verification results include a binary judgment of the truth or falsehood of the factual claims, a confidence quantification of the binary judgment of truth or falsehood, source information of the argument materials, and verification conclusions of logical relevance.

[0013] As a preferred technical solution, the ruling in S3 includes a ruling instruction and a ruling reason; the ruling instruction includes an acceptance instruction, a rejection instruction, or a replanning instruction; the ruling reason adopts a four-part argumentation structure of problem identification, evidence citation, standard comparison, and conclusion deduction; when the ruling instruction is a rejection, a revised draft is generated simultaneously based on the current ruling reason and output together with the ruling result.

[0014] As a preferred technical solution, in process S3, a mechanism for counting the number of review rounds and judging thresholds is introduced: after obtaining the ruling result, it is determined whether the current number of review rounds is greater than the preset maximum number of review rounds. If so, the current draft answer is taken as the final draft and step S4 is executed; otherwise, the process in step S3 is continued to be executed to judge the specific content in the ruling result. The initial number of review rounds is one, and the number of rounds is only increased when the ruling result is a rejection instruction or a replanning instruction.

[0015] According to another aspect of the present invention, a large-scale model answer generation system based on factual reasoning is provided, the system comprising a router module, a draft generation module, a deliberation module, and a stylist module; The router module performs an internal complexity analysis on the input problem and generates problem type information based on the analysis results, classifying the problem as a simple problem or a complex problem. The draft generation module's input end is connected to the router module and the review module, and its output end is connected to the review module; the draft generation module is used to generate and output a draft answer based on the received questions and question type information; The input end of the review module is connected to the draft generation module, and the output end is connected to the draft generation module and the stylist module. It is used to complete the cyclical review process. If a draft fails to pass the review, it is output to the draft generation module for modification or regeneration. If a draft passes the review, it is output as the final draft to the stylist module. The stylist module is used to optimize the final draft and output the result as the answer text.

[0016] As a preferred technical solution, the draft generation module includes a first submodule and a second submodule. The first submodule and the second submodule are in a logically parallel architecture, and one of them is selected to be started according to the problem type: when the problem is a complex problem, the first submodule is started to generate the answer draft; when the problem is a simple problem, the second submodule is started to generate the answer draft. The first submodule includes a planner unit, a multiple proposer unit and a synthesizer unit connected in sequence, and the second submodule includes a draft generation unit. The planner unit develops a solution outline for the input problem; The multi-proposer unit includes multiple large language model instances. Based on the solution outline, the multi-proposer unit calls multiple large language model instances in parallel. Each instance receives the same solution outline and is configured with different perspective instructions to generate multiple drafts from different perspectives. The synthesizer unit integrates all the drafts generated by the multiple proposer units to generate a consensus draft as the answer draft; The draft generation unit directly generates a draft answer to the input question.

[0017] As a preferred technical solution, the deliberation module includes a skeptical unit, a researcher unit, and a judge unit that interact sequentially, and the output of the judge unit is connected to the draft generation module: the questioning information output by the skeptical unit is transmitted to the researcher unit, the evidence report generated by the researcher unit and the questioning information of the skeptical unit are transmitted together to the judge unit, and the judge unit outputs the ruling result based on the questioning information of the skeptical unit and the evidence report of the researcher unit.

[0018] Compared with the prior art, the present invention has the following beneficial effects: 1. In this invention, after generating a draft answer, a review process is conducted. First, the draft answer is challenged. Then, based on the challenge, fact-checking is performed and evidence is provided. Finally, based on the evidence and the challenge, a ruling is made on the draft answer, resulting in a ruling. If the ruling includes an acceptance instruction, the current draft answer is taken as the final draft. If the ruling includes a rejection instruction, the current draft answer is modified and the review process is repeated. If the ruling includes a replanning instruction, a new draft answer is generated and the review process is repeated. This iterative review cycle mechanism overcomes the deficiency of lacking process-based self-correction capabilities in existing technologies, effectively suppresses model illusions, and significantly improves the factual accuracy and logical rigor of the output content.

[0019] 2. In this invention, by quantifying reasoning jumps and evaluating complexity, the fuzzy manual classification logic is replaced, making the criteria for determining problem types quantifiable and reproducible. This avoids solution mismatch due to classification bias. Furthermore, based on complexity, it distinguishes between simple and complex problems, ensuring that simple problems generate answers quickly and complex problems obtain sufficient analytical resources, thereby improving the execution accuracy and resource utilization efficiency of the entire method.

[0020] 3. In this invention, a draft generation mechanism is constructed for complex problems, which involves multi-perspective planning, parallel generation, and consensus synthesis. It covers the blind spots in problem analysis through multiple perspectives such as neutrality, criticality, and quantification. At the same time, it integrates consensus content and unique information to enrich the draft answer, which not only ensures the comprehensiveness of the answer but also avoids the one-sidedness of a single perspective, making the answer to complex problems more in-depth and objective.

[0021] 4. This invention achieves precise and targeted identification of draft answer defects by comprehensively covering four types of challenges: factual, logical, causal, and adversarial. Combined with a priority labeling mechanism, it prioritizes core errors, avoids wasting resources on ineffective challenges, improves the relevance of subsequent fact-checking and draft revision, and shortens the review cycle. Simultaneously, by proactively and structurally raising verifiable challenges to the draft's statements and logical chains, it overcomes the shortcomings of opaque model reasoning processes and difficulty in locating defects in existing technologies. This brings the advantages of dynamic detection of reasoning capabilities and process transparency, accurately exposing the model's shortcomings in the process of solving specific problems.

[0022] 5. In this invention, fact-checking is performed based on the capabilities of the large model itself and in collaboration with external APIs, generating a multi-dimensional evidence report that includes authenticity determination, confidence level, source information, and logical verification. This provides solid factual support for the draft ruling, overcoming the limitations of a single verification method. Furthermore, by tracing sources and quantifying confidence, the authoritativeness and verifiability of the evidence are ensured, providing a clear basis for subsequent amendments. Moreover, this invention employs a four-part structured ruling reasoning, making the defect determination and ruling logic traceable and unambiguous. Simultaneously, a revised draft is generated under a rejection instruction, avoiding the inefficiency of merely identifying defects without providing a direction for rectification. This ensures the targeted nature of the draft amendment, reduces the number of repeated reviews, improves iterative optimization efficiency, and guarantees the factual accuracy and logical coherence of the revised draft.

[0023] 6. In this invention, by introducing a mechanism for counting the number of review rounds and judging thresholds, the quality of the draft is continuously optimized through multiple rounds of iterative correction, while avoiding the efficiency loss caused by infinitely repeated reviews, thus achieving a balance between quality and efficiency. At the same time, the rules for increasing the number of rounds are clearly defined, making the review process controllable and predictable, and improving the feasibility of the method for engineering implementation.

[0024] 7. In this invention, the functional decomposition and collaborative linkage of four major modules—router, draft generation, deliberation, and stylist—achieves modular execution of problem classification, draft generation, cyclical deliberation, and final draft optimization. Each module has clearly defined responsibilities and low coupling, facilitating individual upgrades and optimizations while ensuring the overall stability and scalability of the system, adapting to answer generation needs in different scenarios. The draft generation module is divided into logically parallel complex problem sub-modules and simple problem sub-modules. Combined with the refined division of labor among the planner, multiple proposers, and synthesizer, customized generation of problems of varying complexity is achieved: complex problems are ensured comprehensiveness through multi-unit collaboration, while simple problems are improved efficiency through direct generation. Simultaneously, the clear module architecture facilitates technology implementation and subsequent functional expansion. Through the sequential interaction and feedback loop of the skeptic, researcher, and judge units, the skeptic unit accurately identifies defects, the researcher unit provides authoritative evidence, and the judge unit makes scientific rulings and provides feedback. This ensures the professionalism and standardization of the deliberation process and, through the signal loop between units, achieves full-process control from defect identification to correction, improving the stability of answer quality.

[0025] This invention constructs four major modules—router, draft generation, deliberation, and stylist—and several units under each module, forming a multi-agent collaborative framework based on role division of labor. This framework decomposes complex reasoning tasks into a series of clearly defined sub-tasks, addressing the shortcomings of existing technologies in dealing with complex and open-ended problems. It also enhances the ability to perform structured analysis of complex problems, resulting in more in-depth, comprehensive, and robust final output answers. Attached Figure Description

[0026] Figure 1 This is a schematic diagram illustrating the steps of a large-scale model answer generation method based on factual reasoning in this invention; Figure 2 This is a schematic diagram illustrating the specific process of performing intrinsic complexity analysis on the input problem in this invention; Figure 3 This is a schematic diagram illustrating the specific process of multi-perspective planning, generation, and synthesis to obtain a draft answer in this invention. Figure 4 This is a flowchart illustrating the process of generating answers for a large model based on factual reasoning, as shown in the example. Detailed Implementation

[0027] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0028] Example 1 In this embodiment, a large-scale model answer generation method based on factual reasoning is adopted, and the method steps are as follows: Figure 1 As shown, it specifically includes: S1. Perform an internal complexity analysis on the input problem and classify the problem as a simple problem or a complex problem based on the analysis results; S2. If the current problem is a simple problem, then generate a draft answer directly for the input problem and proceed to step S3; if the current problem is a complex problem, then plan, generate and synthesize a draft answer from multiple perspectives for the input problem. S3. Raise questions about the draft answer, conduct fact-checking based on the questions and provide evidence, and finally make a ruling on the draft answer based on the evidence and questions to obtain a ruling result. If the ruling result includes an acceptance instruction, the current draft answer is taken as the final draft and step S4 is executed; if the ruling result includes a rejection instruction, the current draft answer is modified and step S3 is executed again; if the ruling result includes a replanning instruction, the process jumps to step S2. S4. Optimize the final draft and output it as the answer text.

[0029] The specific process of performing intrinsic complexity analysis on the input problem in S1 is as follows: Figure 2 As shown, it includes: Preset inference jump threshold; Call a large language model instance, use the preset inference hop count evaluation instruction to break down the problem into inference steps, and count the number of inference hops required for the problem; If the number of reasoning hops required for a problem is less than or equal to a preset reasoning hops threshold, the problem is classified as a simple problem; otherwise, it is classified as a complex problem.

[0030] The specific process of multi-perspective planning, generation, and synthesis to obtain a draft answer in S2 is as follows: Figure 3 As shown, it includes: For the input problem, a structured solution outline is developed by combining domain knowledge and problem analysis dimensions; Based on the solution outline, multiple large language model instances are invoked in parallel. Each instance receives the same solution outline and is configured with different perspective instructions, thereby generating multiple drafts from multiple perspectives. Receive all generated drafts, identify them through text semantic analysis, and integrate all drafts into a consensus draft as the answer draft based on the consensus content, differences and unique information among the drafts; Among them, the multiple perspectives include a neutral perspective, a critical perspective that focuses on risk, and a quantitative perspective that focuses on data.

[0031] When implementing the system, such as Figure 4 As shown, it includes: 1. Problem Classification (Route Determination) After the user enters a question, the process first enters the routing stage, where the question is divided into complex queries and simple queries based on its inherent complexity. This is the starting point for the process branches.

[0032] 2. Draft generation branch Complex query branch: If the problem is determined to be complex, the process enters the sub-process of: planner formulating plan, multi-perspective draft generation and forming consensus draft.

[0033] The planner develops a plan: to address the multi-dimensional needs of complex problems, it first develops a structured solution outline; Multi-perspective draft generation: Based on the outline, multiple drafts are generated in parallel from different perspectives (such as neutral, critical, quantitative, etc.); Forming a consensus draft: Integrating the consensus, differences, and unique information from multiple drafts to generate a logically coherent consensus draft.

[0034] Simple query branch: If the problem is determined to be simple, an initial draft is quickly generated and optimized in subsequent steps.

[0035] 3. Deliberation cycle Both the consensus draft of a complex branch and the initial draft of a simple branch need to enter the review cycle: Raise questions: Generate high-priority questions regarding the facts, logic, causality, and other dimensions of the draft; Evidence report generation: Based on the challenge, the system calls a large model or external API to perform fact-checking and outputs an evidence report that includes the determination of authenticity, source information, and logical verification. Overall assessment: Based on the questions raised and the evidence report, a decision will be made to accept, reject, or reconsider the draft: If the instruction is "accept", the draft proceeds to the next stage; if the instruction is "reject", the draft is revised and re-enters the review cycle; if the instruction is "replan", it reverts to the planner's planning stage and a new draft is generated.

[0036] 4. Final output (polished final draft) The draft, after going through a review cycle, eventually enters the polishing and finalization stage, where language expression, logical fluency, and other aspects are optimized to ultimately output a factually accurate and logically rigorous answer text.

[0037] The entire architecture addresses the issues of "factual illusions and logical loopholes" that are prone to occur in large models through the design of problem classification, multi-branch draft generation, iterative review cycle, and final draft optimization. It achieves the accuracy and traceability of answer generation, and each correction is based on clear questioning and evidence links to ensure the reliability of the output answer.

[0038] Example 2 In this embodiment, a large-scale model answer generation system based on factual reasoning is adopted. The system includes a router module, a draft generation module, a deliberation module, and a stylist module. The router module performs an internal complexity analysis on the input problem and generates problem type information based on the analysis results, classifying the problem as a simple problem or a complex problem. The draft generation module's input end is connected to the router module and the review module, and its output end is connected to the review module; the draft generation module is used to generate and output a draft answer based on the received questions and question type information; The input end of the review module is connected to the draft generation module, and the output end is connected to the draft generation module and the stylist module. It is used to complete the cyclical review process. If a draft fails to pass the review, it is output to the draft generation module for modification or regeneration. If a draft passes the review, it is output as the final draft to the stylist module. The stylist module is used to optimize the final draft and output the result as the answer text.

[0039] The draft generation module includes a first submodule and a second submodule. The first and second submodules are logically parallel and are selected to be started according to the problem type: when the problem is complex, the first submodule is started to generate the draft answer; when the problem is simple, the second submodule is started to generate the draft answer. The first submodule includes a planner unit, a multiple proposer unit and a synthesizer unit connected in sequence, and the second submodule includes the draft generation unit. The planner unit develops a solution outline for the input problem; The multiple proposer unit includes multiple large language model instances. Based on the solution outline, the multiple proposer unit calls multiple large language model instances in parallel. Each instance receives the same solution outline and is configured with different perspective instructions to generate multiple drafts from different perspectives. The synthesizer unit integrates all the drafts generated by the multiple proposer units to generate a consensus draft as the answer draft; The draft generation unit directly generates a draft answer to the input question.

[0040] The deliberation module includes a skeptical unit, a researcher unit, and a judge unit that interact sequentially. The output of the judge unit is connected to the draft generation module: the questioning information output by the skeptical unit is transmitted to the researcher unit, the evidence report generated by the researcher unit and the questioning information from the skeptical unit are transmitted to the judge unit, and the judge unit outputs the ruling based on the questioning information from the skeptical unit and the evidence report from the researcher unit.

[0041] 1) Query Analysis and Path Decision: After receiving a user query, the inherent complexity of the query is first analyzed to determine whether it is a simple question that can be answered directly or a complex question that requires in-depth analysis, and different processing paths are selected accordingly.

[0042] 2) Structured Planning and Multi-Perspective Draft Generation: For complex problems, the planning agent first formulates a structured, step-by-step solution outline. Then, multiple proposing agents generate diverse initial drafts from different perspectives based on this outline, and a synthetic agent integrates them into a comprehensive consensus draft.

[0043] 3) Iterative Review and Revision: The consensus draft is placed within a core review cycle. Within this cycle, the skeptical unit is responsible for raising rigorous and verifiable questions about the draft; the researcher agent is responsible for fact-checking and providing evidence to address these questions; and finally, the judge agent synthesizes all information to make a ruling. The ruling may result in acceptance of the draft, rejection and revision for the next round of review, or triggering a redesign due to the discovery of fundamental problems.

[0044] 4) Final Draft Polishing and Output: After the draft passes the review cycle, the stylist module optimizes and formats the text linguistically, and finally generates and outputs a high-quality and highly reliable answer.

[0045] The router module performs query analysis and path decision-making, aiming to improve overall efficiency by matching optimal resources and processing flows to different types of problems. This is achieved by utilizing a large language model instance. Upon receiving a user's original query, the model evaluates the required "reasoning hops" based on a specific prompt. For example, a direct fact lookup is considered 1 hop, a problem requiring one step of deduction is 2 hops, and a complex problem requiring comprehensive analysis is multiple hops. Based on a preset threshold (e.g., ≤2 hops), the router module ultimately outputs a JSON object to decide whether to use a simplified "fast path" or a more in-depth "full deliberation path."

[0046] The draft generation module performs structured planning and multi-perspective draft generation. Its purpose is to construct a clear analytical framework for complex problems and ensure comprehensive content through multi-perspective input, thus avoiding cognitive biases from a single model. The implementation is as follows: First, the planner unit receives complex queries and outputs an execution plan in Markdown format containing multiple logical steps. Next, the multiple proposer unit calls multiple large language model instances in parallel. Each instance receives the same plan but is appended with different perspective instructions (such as "please write from a critical perspective" or "please focus on data and quantitative analysis"), thereby generating multiple complementary drafts. Finally, the synthesizer unit receives all drafts and, by identifying consensus, differences, and unique information, integrates them into a logically coherent and content-rich consensus draft.

[0047] The review module performs iterative review and revision, simulating an adversarial academic peer review process. During answer generation, it proactively exposes and corrects factual and logical flaws, thereby dynamically detecting the model's reasoning ability. This is implemented as a loop, with each loop consisting of the following steps: First, the skeptic unit receives the current draft and outputs a structured JSON array. Each object contains a challenge to a specific statement in the draft, categorized as factual, logical, causal, or adversarial, and assigned high, medium, and low priorities—the core of the detection function. Next, for the high-priority questions raised by the skeptic unit, the researcher unit calls the large language model itself or an external API to perform fact-checking and generates an evidence report containing the results. Finally, the judge unit receives the current draft and the evidence report, performs a comprehensive evaluation, and outputs a JSON object containing a result instruction (accept, reject, or rewrite), the reasoning for the decision, and a potentially revised new draft as the decision result. If the result instruction is rejection, the revised new draft will serve as input for the next round of review, re-entering the loop.

[0048] The stylist module handles the final polishing and output, aiming to improve the readability and professionalism of the final answer. This is achieved by receiving the approved final draft and guiding a large language model with specific instructions to optimize the text's language expression, grammatical structure, and formatting without altering any facts or logic.

[0049] In this embodiment, the specific implementation process of the system is as follows: Step 1: Query Analysis and Path Selection. The system receives a user query: "Please comprehensively analyze the advantages and disadvantages of developing nuclear energy in Country D and provide a conclusion." The router module is invoked. Its internal large language model receives the query and analyzes it according to preset instructions. It determines that the problem involves multi-faceted comparative analysis and is therefore a complex problem. Therefore, it outputs a JSON object: { "decision": "FULL_MAVEN", "estimated_hops": ">2", "reason": "The query requires a cost-benefit analysis and a comprehensive conclusion, necessitating multiple steps of reasoning and information integration." } The system selects the complete review path based on the decision field.

[0050] Step two: Structured planning and multi-perspective draft generation. The planner module is activated. It outputs an execution plan in Markdown format, stored in the central workspace: # Analysis of the advantages and disadvantages of developing nuclear energy 1. **Introduction**: Briefly describe the definition of nuclear energy and its position in country D's energy strategy.

[0051] 2. **Advantages:** This section elaborates on the advantages in energy security, environmental protection (carbon neutrality), and technological autonomy.

[0052] 3. **Disadvantages**: This section will discuss the risks associated with nuclear safety, nuclear waste disposal, and construction costs and timelines.

[0053] 4. **Conclusion:** Considering both advantages and disadvantages, a balanced conclusion and future outlook are presented.

[0054] The multi-proposer module will launch three instances in parallel, generating drafts from different perspectives: Draft A (neutral perspective), Draft B (critical perspective focusing on risk), and Draft C (quantitative perspective focusing on economy and data). Subsequently, the synthesizer module will integrate these three drafts to form a more comprehensive "Consensus Draft 1.0".

[0055] Step 3: Iterative Review and Revision. Draft Consensus 1.0 enters the review cycle (iteration = 1). First, the Skeptic module reviews Draft 1.0 and finds the sentence: "Country D's fourth-generation nuclear power technology is fully mature and can ensure absolute safety." Based on this, the module generates a JSON object and stores it in the workspace. { "claim": "Country D's fourth-generation nuclear power technology is fully mature and can ensure absolute safety." Question: Please verify the current stage of development of 'fourth-generation nuclear power technology' in Country D (e.g., whether it has been deployed on a large scale commercially) and assess the accuracy of the claim of 'absolute safety,' and whether there are any known risks or challenges? "type": "Factual", "priority": "High" } The researcher module detected this high-priority issue, invoked the large model knowledge base for internal fact-checking, and generated an evidence report: "Country D's fourth-generation nuclear power technology (such as high-temperature gas-cooled reactors) has completed demonstration projects and connected to the grid for power generation, demonstrating leading technology, but it has not yet entered the stage of large-scale commercial deployment. In engineering, there is no concept of 'absolute safety.' Although its safety design (such as passive safety systems) has greatly reduced the risk of accidents, there are still potential risks such as extreme natural disasters and human error." The judge module receives the consensus draft 1.0 and the evidence report. Upon comparison, it finds that the statements "fully mature" and "absolutely safe" in the draft are inaccurate and constitute exaggerations. The judge's internal LLM temperature parameter is set to 0.0 to ensure the certainty of the decision. It outputs the ruling in JSON format. { "Reasoning": "The terms 'fully mature' and 'absolutely safe' in the draft are inconsistent with the facts in the evidence report and constitute exaggerated statements that need to be revised. They should be replaced with more rigorous and accurate descriptions." "revised_draft": "...Country D's fourth-generation nuclear power technology has achieved world-leading breakthroughs, and its demonstration project has been successfully connected to the grid. Its advanced passive safety design has greatly improved safety, but it is still far from large-scale commercial application, and no engineering system can promise 'absolute safety'...", "final_decision": "REJECT" } The system detected the decision as REJECT, designated the revised draft as the new "Consensus Draft 2.0," and began the second review cycle (iteration = 2). In the second cycle, the skeptic module did not find any new high-priority factual errors. After evaluation, the judge module determined that Draft 2.0 was consistent with known facts and logically sound, and output the following decision: { "final_decision": "ACCEPT", ... } The deliberation cycle ends. The maximum number of iterations (MAX_ITERATIONS) for the entire cycle is set to 2 in this embodiment.

[0056] Step four, final polishing and output. The stylist module receives the final approved draft, optimizes its paragraphs and polishes its sentences to create a clear, fluent final version, which is then output to the user.

[0057] In this embodiment, the key parameter configurations for each module can be found in Table 1.

[0058] Table 1. Partial Training Parameters In summary, this solution, by adopting an iterative review cycle mechanism, which includes a closed-loop process of doubt, verification, and adjudication, solves the deficiency of lack of process-based self-correction capability caused by the one-time generation mode in the existing technology. It brings the advantages of significantly improving the factual accuracy and logical rigor of the answer, thereby effectively suppressing the generation of model illusion.

[0059] Secondly, by setting up a dedicated skeptic unit, whose task is to proactively and structurally raise verifiable questions about the draft's statements and logical chains, the shortcomings of the existing technology, such as the lack of transparency in the model's reasoning process and the difficulty in locating defects, are resolved. This brings the advantages of achieving dynamic detection of reasoning ability and transparency of the process, and can accurately expose the model's shortcomings in the process of solving specific problems.

[0060] Finally, by constructing a multi-agent collaborative framework (planning, proposing, deliberating, etc.) based on role division of labor, complex reasoning tasks are decomposed into a series of well-defined sub-tasks. This solves the shortcomings of existing technologies in dealing with complex and open problems, and brings the advantage of enhancing the ability to perform structured analysis of complex problems, making the final output answer more in-depth, comprehensive and robust.

[0061] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered 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 method for generating answers for large-scale models based on factual reasoning, characterized in that, The method steps include: S1. Perform an internal complexity analysis on the input problem and classify the problem as a simple problem or a complex problem based on the analysis results; S2. If the current problem is a simple problem, then generate a draft answer directly for the input problem and proceed to step S3; if the current problem is a complex problem, then plan, generate and synthesize a draft answer from multiple perspectives for the input problem. S3. Raise questions about the draft answer, conduct fact-checking based on the questions and provide evidence, and finally make a ruling on the draft answer based on the evidence and questions to obtain a ruling result. If the ruling result includes an acceptance instruction, the current draft answer is taken as the final draft and step S4 is executed; if the ruling result includes a rejection instruction, the current draft answer is modified and step S3 is executed again; if the ruling result includes a replanning instruction, the process jumps to step S2. S4. Optimize the final draft and output it as the answer text.

2. The method for generating large-scale model answers based on factual reasoning according to claim 1, characterized in that, The specific process of performing intrinsic complexity analysis on the input problem in S1 includes: Preset inference jump threshold; Call a large language model instance, use the preset inference hop count evaluation instruction to break down the problem into inference steps, and count the number of inference hops required for the problem; If the number of reasoning hops required for a problem is less than or equal to a preset reasoning hops threshold, the problem is classified as a simple problem; otherwise, it is classified as a complex problem.

3. The method for generating large-scale model answers based on factual reasoning according to claim 1, characterized in that, The specific process of multi-perspective planning, generation, and synthesis to obtain the draft answer in S2 includes: For the input problem, a structured solution outline is developed by combining domain knowledge and problem analysis dimensions; Based on the solution outline, multiple large language model instances are invoked in parallel. Each instance receives the same solution outline and is configured with different perspective instructions, thereby generating multiple drafts from multiple perspectives. Receive all generated drafts, identify them through text semantic analysis, and integrate all drafts into a consensus draft as the answer draft based on the consensus content, differences and unique information among the drafts; Among them, the multiple perspectives include a neutral perspective, a critical perspective that focuses on risk, and a quantitative perspective that focuses on data.

4. The method for generating large-scale model answers based on factual reasoning according to claim 1, characterized in that, When questioning the draft answer in S3, the questions raised include factual questions, logical questions, causal questions, and adversarial questions, and the questions are marked with priority.

5. The method for generating large-scale model answers based on factual reasoning according to claim 4, characterized in that, When performing fact-checking and providing evidence based on the challenge in S3, the specific steps are as follows: calling the large language model itself or an external API to perform fact-checking and generating an evidence report containing the verification results; the verification results include a binary determination of the truth or falsehood of the factual claims, a confidence quantification of the binary determination of truth or falsehood, source information of the argument materials, and verification conclusions of logical relevance.

6. The method for generating large-scale model answers based on factual reasoning according to claim 5, characterized in that, The ruling in S3 includes a ruling instruction and a ruling reason; the ruling instruction includes an acceptance instruction, a rejection instruction, or a replanning instruction; the ruling reason adopts a four-part argumentation structure of problem identification, evidence citation, standard comparison, and conclusion deduction; when the ruling instruction is a rejection, a revised draft is generated simultaneously based on the current ruling reason and output together with the ruling result.

7. The method for generating large-scale model answers based on factual reasoning according to claim 1, characterized in that, In the S3 process, a mechanism for counting the number of review rounds and judging thresholds is introduced: after obtaining the ruling result, it is determined whether the current number of review rounds is greater than the preset maximum number of review rounds. If so, the current draft answer is taken as the final draft and step S4 is executed; otherwise, the process in step S3 is continued to be executed to judge the specific content in the ruling result. The initial number of review rounds is one, and the number of rounds is only increased when the ruling result is a rejection instruction or a replanning instruction.

8. A large-scale model answer generation system based on factual reasoning, characterized in that, The system operates using a large model answer generation method based on factual reasoning as described in any one of claims 1-7. The system includes a router module, a draft generation module, a deliberation module, and a stylist module. The router module performs an internal complexity analysis on the input problem and generates problem type information based on the analysis results, that is, classifies the problem into a simple problem or a complex problem. The draft generation module has its input end connected to the router module and the review module, and its output end connected to the review module. The draft generation module is used to generate and output a draft answer based on the received question and question type information; The input end of the review module is connected to the draft generation module, and the output end is connected to the draft generation module and the stylist module. It is used to complete the cyclic review process. If a draft fails to pass the review, it is output to the draft generation module for modification or regeneration. If a draft is approved, it will be output as the final draft to the stylist module; The stylist module is used to optimize the final draft and output the answer text.

9. A large-scale model answer generation system based on factual reasoning according to claim 8, characterized in that, The draft generation module includes a first submodule and a second submodule. The first submodule and the second submodule are in a logically parallel architecture, and one of them is selected to be started according to the problem type: when the problem is a complex problem, the first submodule is started to generate a draft answer; when the problem is a simple problem, the second submodule is started to generate a draft answer. The first submodule includes a planner unit, a multiple proposer unit and a synthesizer unit connected in sequence, and the second submodule includes a draft generation unit. The planner unit develops a solution outline for the input problem; The multiple proposer unit includes multiple large language model instances. Based on the solution outline, the multiple proposer unit calls multiple large language model instances in parallel. Each instance receives the same solution outline and is configured with different perspective instructions to generate multiple drafts from different perspectives. The synthesizer unit integrates all the drafts generated by the multiple proposer units to generate a consensus draft as the answer draft; The draft generation unit directly generates a draft answer to the input question.

10. A large-scale model answer generation system based on factual reasoning according to claim 8, characterized in that, The deliberation module includes a skeptical unit, a researcher unit, and a judge unit that interact sequentially. The output of the judge unit is connected to the draft generation module: the questioning information output by the skeptical unit is transmitted to the researcher unit, the evidence report generated by the researcher unit and the questioning information of the skeptical unit are transmitted to the judge unit, and the judge unit outputs the ruling based on the questioning information of the skeptical unit and the evidence report of the researcher unit.

Images