A knowledge exploration method and system based on generative thinking chain and feedback mechanism

Abstract

The application discloses a kind of knowledge exploration method and system based on generative thinking chain and feedback mechanism, the method includes: receiving the multimodal query of user input, carries out intention analysis and identifies core demand;Multimodal query pre-processing and extracting key information, complex problem is decomposed into multiple sub-problems or sub-tasks;Based on generative thinking chain dynamic construction layered reasoning path, label logic basis for each step reasoning path;Through dynamic weight distribution optimization reasoning path priority, reasoning strategy is combined with meta reinforcement learning framework and adjusted online dynamically, and the reasoning path is optimized;Using causal reasoning analysis tool to check reasoning path, generate counterfactual path to correct error node;Output structured text and visual reasoning process, and based on game theory, the model parameter is optimized in real time in the feedback aggregation of multiple people.The application can effectively solve the intelligent query and reasoning problem in complex scene, realize the precision, real-time and trust of knowledge service.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence (AI), specifically to the cross-technology field of intelligent query and knowledge reasoning technology, and more particularly to a knowledge exploration method and system based on generative thinking chain and feedback mechanism. BACKGROUND

[0002] Existing intelligent query systems mostly use static rules or single modal input, rely on keyword matching and preset reasoning templates, and have the following limitations:

[0003] 1. Traditional systems cannot dynamically adjust reasoning paths according to context, and when faced with complex problems (such as legal text conflicts, multi-modal input fusion), the retrieval efficiency is low and the results are single.

[0004] Existing AI models often require large computing resources and time investment when dealing with long text and high complexity query tasks. Due to the large amount of data, the traditional reasoning process is slow, especially when facing high concurrency query requests, the response speed of the system decreases significantly, resulting in poor user experience.

[0005] In complex situations such as legal regulations or academic papers, existing models are prone to problems such as inaccurate key information extraction, logical reasoning errors, or context understanding biases. In this case, the returned results may not be accurate or relevant, making it difficult to meet users' demand for high-quality knowledge services. Existing technologies are difficult to capture deep semantic associations in long texts or multi-round interactions, especially in professional fields such as law and medicine, making it difficult to accurately match user intent.

[0006] Most existing technologies use static rules or fixed reasoning path generation strategies, which can easily lead to paths that deviate from expectations during reasoning. This instability affects the accuracy and consistency of knowledge derivation, especially when dealing with variable query requests, making it difficult to ensure stable results.

[0007] 2. Traditional feedback mechanisms only rely on offline user behavior data and cannot optimize reasoning paths through real-time interaction, resulting in limited system response speed and personalization level.

[0008] Existing systems lack a mechanism to adjust reasoning parameters or paths in real time based on feedback after reasoning is complete. As the application environment changes, the model cannot be updated in a timely manner to maintain high performance, and over time, the system performance gradually declines, making it difficult to continuously provide efficient reasoning services.

[0009] Existing models often present "black box" characteristics, and the output results lack detailed reasoning process display. Especially in high-risk fields such as law and medicine, users expect to understand the specific steps and basis of reasoning, but this demand cannot be effectively met in the prior art, reducing the user's trust in the system results.

[0010] Therefore, how to better overcome the limitations of the above intelligent query system has become a problem that technicians in the field need to solve. SUMMARY

[0011] Therefore, the present application provides a knowledge exploration method and system based on generative thinking chain and feedback mechanism, which can effectively solve the intelligent query and reasoning problem in complex scenarios, and realize the precision, real-time and credibility of knowledge service.

[0012] In order to achieve the above purpose, the present application adopts the following technical solutions:

[0013] In a first aspect, the present application provides a knowledge exploration method based on generative thinking chain and feedback mechanism, comprising the following steps:

[0014] S10, receiving a multi-modal query input by a user, identifying core requirements, writing topics or problem focuses through intent analysis;

[0015] S20, preprocessing the multi-modal query and extracting key information, decomposing complex problems into multiple sub-problems or sub-tasks;

[0016] S30, according to the multiple sub-problems or sub-tasks, dynamically constructing a hierarchical reasoning path based on the generative thinking chain, and labeling the logical basis for each step of the reasoning path;

[0017] S40, optimizing the priority of the reasoning path through dynamic weight allocation, combining the meta-reinforcement learning framework to dynamically adjust the reasoning strategy online, and optimizing the reasoning path;

[0018] S50, on the basis of optimizing the reasoning path, using a causal reasoning analysis tool to verify the reasoning path, and generating an alternative path to correct error nodes;

[0019] S60, outputting structured text and visual reasoning process, and based on the multi-person feedback aggregation of game theory, optimizing the model parameters in real time.

[0020] In one embodiment, in the step S40, the priority of the reasoning path is optimized through dynamic weight allocation, comprising:

[0021] According to the task context, semantic similarity and node importance, the weight of each node in the reasoning path is dynamically calculated, and the weight is calculated through the following formula:

[0022]

[0023] wherein w i represents the weight of the reasoning node n i ; q represents the current query task; n i represents the i-th reasoning node; j represents the number of all reasoning nodes in the current path; sim(q, n i ) represents the semantic similarity between the current task q and the reasoning node n i ; and τ represents a temperature coefficient for controlling the path exploration intensity.

[0024] According to the plurality of nodes decomposed from the reasoning path, each node corresponds to a subtask, and the optimal selection of the path is realized by maximizing the total reward value through the reward model; wherein the formula for maximizing the total reward value is as follows:

[0025]

[0026] wherein T represents the query task, R i (T) represents the reward value of the node, k represents the number of nodes in the current path; and p represents a dynamically adjusted normalized weight coefficient, p∈[0,1].

[0027] In one embodiment, the step S40 of optimizing the reasoning path priority through dynamic weight distribution further comprises:

[0028] The multi-modal input information is fused, and the semantic alignment of the text and the image is realized through an alignment loss function, and the loss function is defined as:

[0029]

[0030] wherein II ab ∈{0,1} represents a modal correlation indication function, which is 1 when the text T a is related to the image I b , and 0 otherwise; f T (·) and f I (·) represent a text encoder and an image encoder respectively, and output normalized embedding vectors; I k represents the k-th sample in the image modal, k≠b; cos(·,·) represents a cosine similarity, and τ represents a temperature coefficient; N represents the number of text samples, and M represents the number of image samples.

[0031] In one embodiment, the step S40 of dynamically adjusting the reasoning measurement online through the meta-reinforcement learning framework comprises:

[0032] Offline meta-pretraining, the model is pre-trained on a cross-domain task set by using a MAML framework;

[0033] Online fine-tuning generates counterfactual samples through real-time user feedback to dynamically update model parameters;

[0034] A composite reward mechanism that incorporates explainability indicators balances inference accuracy, efficiency, and explainability through a multi-objective reward function design.

[0035] In one embodiment, the step S50 uses a causal reasoning analysis tool to verify the reasoning path, including:

[0036] Generate counterfactual queries when the reasoning result is biased, compare the actual path with the counterfactual path to locate the error nodes and trigger correction;

[0037] By embedding causal probability edges in knowledge to define the causal strength between nodes, and calculating the logical consistency score, task relevance score, case support score, and explainability score, the reasoning credibility and explainability are enhanced.

[0038] In one embodiment, the step S50 uses a causal reasoning analysis tool to verify the reasoning path, including:

[0039] At each step of the reasoning path generation, the reasoning chain is optimized by a reward model, and the reward mechanism is as follows:

[0040] R=w1·Accuracy+w2·Efficiency-w3·Resources

[0041] w1 represents the weight coefficient of accuracy; Accuracy represents the accuracy score of the reasoning chain; w2 represents the weight coefficient of efficiency; Efficiency represents the efficiency score of the reasoning chain; w3 represents the weight coefficient of resource consumption; Resources represents the resource consumption score of the reasoning chain.

[0042] In one embodiment, the step S60 uses a game theory-based multi-person feedback aggregation to optimize model parameters in real time, including:

[0043] Based on the game theory, the Shapley value algorithm quantifies user feedback contributions, filters low-quality or malicious feedback, and coordinates multiple user opinions to optimize the reasoning path.

[0044] Use the meta-reinforcement learning framework to process the aggregated feedback information in real time, dynamically adjust the reasoning path and model parameters.

[0045] In a second aspect, the embodiments of the present application also provide a knowledge exploration system based on generative thinking chain and feedback mechanism, comprising:

[0046] A data input module for receiving user inputted multi-modal queries, performing intent analysis to identify core requirements, writing topics or problem focuses.

[0047] a semantic processing module configured to preprocess the multi-modal query and extract key information, and decompose a complex problem into multiple sub-problems or sub-tasks;

[0048] a reasoning engine module configured to dynamically construct a hierarchical reasoning path based on a generative thinking chain according to the multiple sub-problems or sub-tasks, and label a logical basis for each step of the reasoning path;

[0049] a dynamic reasoning path module configured to optimize a priority of the reasoning path through dynamic weight distribution, and dynamically adjust a reasoning strategy online in combination with a meta-reinforcement learning framework to optimize the reasoning path;

[0050] a causal reasoning analysis module configured to, on the basis of the optimized reasoning path, check the reasoning path by using a causal reasoning analysis tool, and generate an anti-fact path to correct an error node;

[0051] a result output and feedback optimization module configured to output a structured text and a visual reasoning process, and optimize model parameters in real time based on a multi-person feedback aggregation algorithm of game theory.

[0052] According to the technical solution, compared with the prior art, the application has the following technical advantages:

[0053] Accuracy: through dynamic weight distribution and causal reasoning, the deep logical relationship of the text can be accurately captured, and key information can be avoided.

[0054] Real-time: based on the online optimization framework of meta-reinforcement learning, the dynamic changes of the text can be quickly adapted.

[0055] Reliability: the result is enhanced by causal reasoning and explainability, and meets the strict requirements of high-risk fields.

[0056] Efficiency: the overall architecture design improves the operation efficiency of large-scale reasoning tasks, and is suitable for real-time applications in complex scenarios.

[0057] The application integrates various advanced AI technologies, including a dynamic weight distribution reasoning path generation mechanism, a meta-reinforcement learning, a reasoning path optimization framework based on meta-reinforcement learning, a causal reasoning, a multi-person feedback aggregation algorithm based on game theory, and an explainability enhancement technology, which significantly improves the performance of the intelligent query and knowledge reasoning system, and provides reliable technical support for knowledge services in complex scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0058] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of the provided drawings.

[0059] Figure 1 The knowledge exploration method flowchart based on the generative thinking chain and feedback mechanism provided by the present application.

[0060] Figure 2 The knowledge exploration technology roadmap based on the generative thinking chain and feedback mechanism provided by the present application.

[0061] Figure 3 The knowledge exploration system architecture diagram based on the generative thinking chain and feedback mechanism provided by the present application. DETAILED DESCRIPTION

[0062] The technical solutions in the embodiments of the present application will be described clearly and completely in the following with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0063] The complex scenarios targeted by the present application involve a class of texts with the following characteristics:

[0064] Formal characteristics: formal and rigorous long texts, complex structure, containing multi-level logical relationships; the text is dense in terms, with high semantic ambiguity, which puts high requirements on information extraction and reasoning; the text content is frequently updated dynamically, requiring the system to have efficient real-time updating and adaptive capacity.

[0065] Application characteristics: strong multi-dimensional correlation, reasoning needs to integrate multiple information sources; the scenario involves high-risk decision-making, requiring the reasoning result to have extremely high accuracy and transparency; different users have different interpretation needs for the same text, personalized reasoning services are needed; some scenarios require multidisciplinary collaborative reasoning to handle multi-source, multi-modal information.

[0066] These characteristics are embodied in formal and rigorous long texts such as legal regulations, academic papers, technical specifications / standard documents, government reports, contract documents, medical guidelines, financial reports, and user manuals. The core uses include:

[0067] Legal regulations: quickly locate key information in legal clauses to assist professionals in compliance analysis or legal consultation.

[0068] Academic Research & Technical Documentation: Supports knowledge mining in academic papers and literature review generation, accelerating technical researchers' understanding and referencing of technical specifications.

[0069] Business & Financial Analysis: Analyzes financial reports and contract documents, assisting in identifying risk clauses or performing financial data analysis.

[0070] Medical Health: Interprets medical guidelines and case materials, providing references for clinical decision-making.

[0071] Policy Interpretation: Processes government reports and social policy documents, extracting policy highlights and implementation suggestions.

[0072] In modern society, legal regulations, academic papers, technical specifications / standards documents, government reports, contract documents, medical guidelines, financial reports, and user manuals are all formal and rigorous texts that are lengthy and complex. These texts have the following notable characteristics and rules:

[0073] Large-scale texts with complex structures: Texts in these scenarios are often lengthy and have complex internal structures, containing multiple levels of logical relationships, such as progression, parallelism, and conditional association between clauses.

[0074] Terminology-intensive and high semantic ambiguity: Formal texts often contain a large number of professional terms, industry abbreviations, and ambiguous expressions, which increase the difficulty of information extraction and logical reasoning.

[0075] Strong multi-dimensional relevance: Different paragraphs or chapters in the text often have high relevance, and a single conclusion may require reasoning based on information from multiple sources.

[0076] Frequent dynamic updates and rapid version iterations: Legal regulations, technical standards, and other texts are often frequently revised due to policy adjustments or technological innovations, which puts high demands on the real-time updating capabilities of knowledge bases.

[0077] Application scenarios involve high-risk decision-making: In the fields of finance, medicine, and law, any reasoning errors may result in significant economic losses or legal disputes, so the accuracy, reliability, and explainability of the system are extremely high.

[0078] Differences in multi-user needs are obvious: Different users may have different interpretations of the same text, such as legal consultants who may focus on the applicable scope of legal clauses, and technical personnel who may focus on the specific execution steps of operation specifications.

[0079] Strong cross-disciplinary collaboration: For example, the interpretation of medical guidelines may require the combination of clinical experience, pharmacological knowledge, and individual patient circumstances. This multi-disciplinary cross-characteristic makes simple keyword matching insufficient, and deep reasoning capabilities are required.

[0080] In summary, the regularity of these scenarios is that they pose all-round challenges to the knowledge integration capability, reasoning accuracy, dynamic adaptation capability and multi-dimensional feedback mechanism of the knowledge service system. In view of these characteristics, the application adopts the generative thinking chain and dynamic feedback optimization technology, which can effectively solve the intelligent query and reasoning problems in complex scenarios, and realize the precision, real-time and credibility of knowledge service.

[0081] Referring to Figure 1 The embodiment of the application discloses a knowledge exploration method based on a generative thinking chain and a feedback mechanism, comprising the following steps S10-S60:

[0082] S10, receiving a multi-modal query input by a user, performing intent analysis to identify a core requirement, a writing theme or a problem focus;

[0083] The method can be realized by a software program, such as an interactive query platform system, which provides a friendly user interface and supports users to input problems or requirements in natural language. The user first inputs the writing requirement through the system interface. These requirements can be specific writing instructions, natural language descriptions or queries (QA) in the form of a question. After receiving the input, the system immediately performs intent analysis to analyze the user's input and identify the core requirement, writing theme or problem focus, thereby providing a clear direction for subsequent processing.

[0084] S20, preprocessing the multi-modal query and extracting key information, decomposing a complex problem into multiple sub-problems or sub-tasks;

[0085] After completing the intent analysis in step S10, the system preprocesses the Query input by the user, including word segmentation, part-of-speech tagging and other natural language processing operations, to extract key information. Subsequently, the system further extracts key words and key terms in the Query using natural language processing technology. These key words will serve as the basis for generating reasoning paths. For complex problems, the system will also decompose the Query into multiple sub-problems or sub-tasks for subsequent hierarchical reasoning.

[0086] S30, dynamically constructing a hierarchical reasoning path based on the generative thinking chain according to the multiple sub-problems or sub-tasks, and labeling the logical basis for each step of the reasoning path;

[0087] In this step, based on the generative thinking chain technology, the system dynamically constructs a hierarchical reasoning path. The reasoning path starts from the user's input Query and is gradually decomposed into multiple sub-tasks, each corresponding to a node in the reasoning path. The system labels each step of the reasoning path with clear logical basis, such as causal relationship, rule matching or semantic association, which provides a foundation for subsequent weight allocation and optimization. The preliminary reasoning path is presented in a hierarchical structure, ensuring that complex problems can be solved step by step.

[0088] S40, optimize the priority of the reasoning path through dynamic weight allocation, combine the meta-reinforcement learning framework to dynamically adjust the reasoning strategy online, and optimize the reasoning path;

[0089] In this step, the system calculates the weight of each reasoning node through a dynamic weight allocation mechanism. The weight allocation considers task context, semantic similarity and the importance of nodes to ensure that key nodes and important logical relationships are given priority. According to the dynamic weight, the system re-adjusts the priority of the reasoning path to optimize the efficiency and accuracy of the reasoning path. During the reasoning process, the system also dynamically adjusts the weight allocation according to the intermediate results to further optimize the reasoning path.

[0090] The optimized reasoning path then enters the meta-reinforcement learning framework for online optimization. The meta-reinforcement learning framework uses real-time user feedback and historical data to dynamically adjust the reasoning strategy to adapt to new tasks and changing input scenarios. Through policy gradient update and adversarial sample generation, the system continuously optimizes the response speed and accuracy of the reasoning path. At the same time, the meta-reinforcement learning framework combines causal reasoning analysis tools to further strengthen the correctness of the causal relationships in the reasoning path, avoiding errors caused by relevance misjudgment.

[0091] Among them, the dynamic weight allocation mechanism and the meta-reinforcement learning framework are as follows:

[0092] 1. Reasoning path generation mechanism based on dynamic weight allocation

[0093] Traditional thinking chain generation relies on fixed rules or single modal input, while the system realizes real-time optimization and multi-modal collaborative reasoning of the reasoning path through dynamic weight allocation algorithm and cross-modal joint embedding space, which includes:

[0094] Dynamic reasoning node weight calculation:

[0095] An adaptive weight allocation formula based on task context is proposed, which converts the traditional static path selection into a dynamic optimization problem:

[0096]

[0097] w i : represents the reasoning node n ithe weight of the i-th query.

[0098] q: represents the current query task.

[0099] n i : represents the i-th reasoning node.

[0100] j represents the number of all reasoning nodes in the current path;

[0101] sim(q, n i ): represents the semantic similarity between the current task q and the reasoning node n i .

[0102] τ: temperature coefficient, used to control the strength of path exploration.

[0103] Multi-modal joint reasoning path generation:

[0104] Joint embedding method based on cross-modal alignment loss, solves the problem of multi-modal information (text, image, speech) fusion bias.

[0105] The following is an example of text and image, define the alignment loss function:

[0106]

[0107] Where, II ab ∈{0,1} is the modal correlation indicator function, which is 1 when the text T a is related to the image I b , otherwise 0.

[0108] f T (·) and f I (·) are the text encoder and image encoder, respectively, output normalized embedding vectors.

[0109] I k represents the k-th sample in the image modality; k∈(1,M), but k≠b;

[0110] cos(·,·) represents the cosine similarity, and τ is the temperature coefficient (set to 0.07 in the experiment).

[0111] N is the number of text samples, and M is the number of image samples.

[0112] Objective: By maximizing the similarity of related text-image pairs and minimizing the similarity of unrelated pairs, cross-modal semantic alignment is achieved.

[0113] Normalization and temperature coefficient: the embedding vector is L2 normalized, and the temperature coefficient τ controls the sharpness of the similarity distribution, avoiding gradient saturation.

[0114] The system can generate a logical reasoning path and ultimately arrive at the solution to the problem according to the user's input query. The key is to dynamically build the reasoning path, that is, in the context of each problem, the reasoning chain should be automatically generated and adjusted at any time according to the content of the query.

[0115] According to the type of query, complex problems are decomposed into multiple operable sub-tasks or sub-problems. For legal field problems, tasks may include legal text retrieval, responsibility subject analysis, case retrieval, etc.

[0116] According to the type of problem (for example, legal liability, medical problem, business decision, etc.), the system selects the appropriate reasoning framework. These reasoning frameworks include rule-based reasoning, knowledge reasoning.

[0117] The core idea of building a reasoning path is to dynamically select sub-tasks of the reasoning path according to the specific needs of the user query, and adjust the path according to the intermediate results. Assuming that the system has multiple reasoning nodes n1, n2,..., n k , each node represents a reasoning task; k represents the number of nodes in the current path. For a given input query, path selection is determined by the following dynamic weight allocation:

[0118]

[0119] Where T is the query task, W i is the weight of node n i , R i (T) is the reward value of the node (calculated according to the effectiveness and efficiency of the current reasoning result). Path selection is completed by maximizing the total reward of the path:

[0120]

[0121] Where p represents the dynamically adjusted normalized weight coefficient, p ∈ [0, 1].

[0122] 2. Meta-reinforcement learning-based online optimization framework for reasoning path

[0123] In view of the lag problem of the traditional feedback mechanism, a hybrid architecture of meta-reinforcement learning and online supervised learning is proposed to realize real-time correction and long-term stability optimization of the reasoning path:

[0124] Two-stage meta-learning strategy:

[0125] Offline meta-pretraining: using the MAML (Model-Agnostic Meta-Learning) framework, pre-training the model on a cross-domain task set (law, medicine, finance) to enable it to quickly adapt to new tasks.

[0126] Online fine-tuning: Generate adversarial samples through user real-time feedback, dynamically update model parameters, solve domain drift problem.

[0127] Multi-objective reward function design:

[0128] Propose a composite reward mechanism that integrates interpretability indicators:

[0129] R = λ1·Accuracy + λ2·Efficiency + λ3·Interpretability + λ4·UserSatisfaction

[0130] Where λ1, λ2, λ3, λ4 are weight coefficients, Accuracy is accuracy, Efficiency is efficiency, Interpretability is interpretability, and UserSatisfaction represents user satisfaction. The optimal combination is determined through Pareto frontier analysis, satisfying λ1 + λ2 + λ3 + λ4 = 1. The time decay factor r = 0.95 dynamically adjusts the weight of user satisfaction to ensure long-term performance stability of the system.

[0131] For each relationship r, the system generates structured mapping rules by defining its corresponding entity class Ener in the template:

[0132] SchemaNER(r) = {(h, t) | h ∈ HeadEntities(r), t ∈ TailEntities(r)}

[0133] For example, for the relationship r = "responsibility", define:

[0134] SchemaNER("responsibility") = {("company", "employee"), ("legal person", "shareholder")}

[0135] Generate SchemaNERS by task examples to guide the model, ensuring the normativity and consistency of entity prediction.

[0136] To help large language models understand the generation context of named entity patterns, the model can be provided with task background prompts, which contain background information in specific domains, aiming to guide large models to better understand and generate correct entity patterns in the task of law and responsibility classification.

[0137] Task background prompt: The task background prompt provides the generation context for the large model, including the specific definition of the potential relationship r, examples of entity categories Ener, and related legal context. By giving the task background prompt, the system can guide the model to be more accurate when processing entity labeling, reduce irrelevant predictions, and focus on the target task.

[0138] Template filling and data generation: The filling process of the template guides the model to generate the corresponding SchemaNER by giving task examples. For example, given the potential relationship in the task description, such as "the responsibility of the company to the employee", the template will automatically fill in the entity categories and generate a pattern containing "company" and "employee" as head and tail entities, and finally generate a complete named entity pattern SchemaNER.

[0139] The system can ensure that each potential relationship can accurately match the relevant legal provisions and entities when generating the named entity pattern. This typical mapping defines the structured mapping between relationships and entities, ensuring the normativity and consistency of the annotation.

[0140] Example mapping: Assuming that a certain legal provision involves the responsibility of "company" and "employee", the system defines "company" and "employee" as potential entity categories in the named entity template by mapping the relationship "company-responsibility" and "employee-responsibility", helping the subsequent NER model understand their semantic roles.

[0141] Relationship constraints: The system provides clear entity prediction boundaries for the model by limiting the potential relationships r in the task. For example, the task background clearly states that the "responsibility" relationship can only involve the company and employee entities, thereby reducing the prediction of other irrelevant entities.

[0142] Excessive prediction control: Through the mapping control between relationships and entities, the system can ensure that the output of the large model is limited to entities related to the task, avoiding the generation of irrelevant or redundant entity categories. This makes the model more focused and accurate when dealing with complex legal relationships.

[0143] S50, on the basis of optimizing the reasoning path, using causal reasoning analysis tools to check the reasoning path, and generating counterfactual path to correct error nodes;

[0144] On the basis of the optimized reasoning path, the system uses causal reasoning analysis tools to further identify and verify the causal relationships between variables in the reasoning path. Through counterfactual path analysis, the system compares the differences between the actual path and the counterfactual path, locates the error reasoning nodes and triggers correction. At the same time, embed causal probability edges in the knowledge graph, define the causal strength between nodes, and further enhance the explainability and credibility of the reasoning path.

[0145] Among them, the causal-driven reasoning path correction technology is as follows:

[0146] To solve the credibility problem of black box models, introduce causal intervention technology, and build a traceable reasoning chain correction process:

[0147] Counterfactual path analysis:

[0148] When the inference result deviates from the expectation, the system automatically generates a counterfactual question:

[0149] "What would happen if the conclusion of node N_K was ignored?"

[0150] By comparing the actual path with the counterfactual path, the error reasoning node is located and the correction is triggered.

[0151] Causal graph reinforcement:

[0152] Embed causal probability edges in knowledge, define causal strength between nodes: such as this technology enables the system to identify confounding variables and output causal explanations in medical liability determination and other scenarios. By calculating logical consistency score, task relevance score, case support score, and explainability score, to enhance reasoning credibility and explainability.

[0153] Logical consistency score, calculate the logical coherence between each step of the reasoning chain, to ensure that there is no contradictory conclusion before and after.

[0154] Task relevance score, calculate the relevance between user query and generated thought chain through vector similarity. Combined with pre-trained language model, calculate the matching degree of generated content and target problem.

[0155] Case support score, retrieve relevant legal cases from the case library, and calculate the support degree of the current reasoning chain. If the reasoning chain is highly matched with known legal cases, give higher score to enhance the reasoning credibility.

[0156] Explainability score, evaluate whether the reasoning chain provides enough intermediate reasoning steps to make the final conclusion verifiable. If the reasoning path is transparent and easy for manual review, give higher score.

[0157] In order to improve the accuracy and robustness of the reward model, the embodiment adopts reinforcement learning method for optimization. Strategy gradient method is used to optimize the reasoning path of large model, so that it is more consistent with the scoring standard of reward model. Through autoregressive way, multiple candidate reasoning paths are generated, and the path with the highest score is selected as the final output. Combined with artificially annotated high-quality reasoning chain, the reward model is fine-tuned to make it more consistent with the judgment standard of legal experts.

[0158] At each step of the reasoning path generation, the system optimizes the reasoning chain through the reward model. The reward mechanism is as follows:

[0159] R = w1·Accuracy + w2·Efficiency - w3·Resources

[0160] w1: is the weight coefficient of Accuracy. It represents the importance of the accuracy of the inference chain in the reward model. Accuracy generally refers to the degree of consistency between the inference result and the actual result or known correct answer. A higher accuracy weight means that the system tends to choose paths that can provide correct or accurate inference results.

[0161] Accuracy: refers to the accuracy score of the inference chain, which can be calculated by comparing the matching degree of the inference result with the true situation or known cases. In legal reasoning systems, this involves multiple aspects such as logical consistency score, task relevance score, and case support score.

[0162] w2: is the weight coefficient of Efficiency. It represents the importance of the efficiency of the inference process in the reward model. Efficiency generally refers to the time and resources required to complete the inference, and higher efficiency means that the system can generate inference results faster.

[0163] Efficiency: refers to the efficiency score of the inference chain, which can be evaluated by calculating the time required for the inference process, computing resource consumption, etc. In practical applications, improving efficiency can help the system respond to user queries faster, thereby improving user experience.

[0164] w3: is the weight coefficient of Resources. It represents the importance of the negative impact of resource consumption in the inference process in the reward model. Resource consumption generally refers to the computing resources, storage space, etc. required to complete the inference.

[0165] Resources: refers to the resource consumption score of the inference chain, which can be calculated by evaluating the computing resources, storage requirements, etc. used in the inference process. In resource-limited environments, reducing resource consumption can improve the scalability and cost-effectiveness of the system.

[0166] S60, output structured text and visualized inference process, and real-time optimization of model parameters based on game theory multi-person feedback aggregation.

[0167] After the inference path is verified and reinforced by causal reasoning, it finally enters the text generation step. The system generates text content that meets the user's needs based on the optimized inference path and outputs it in a structured form. The generated text not only meets the user's writing needs, but also comes with detailed explanation of the inference process, which is visually displayed through time axis chart, causal relationship diagram, etc. to improve the transparency of the results and the user's trust in the system.

[0168] The user evaluates the generated text content, and the feedback system generates the quality of the results. The system first processes the user's evaluation through a multi-person feedback aggregation algorithm based on game theory, quantifies the feedback contribution of each user, and coordinates the evaluation opinions of different users to ensure the fairness and stability of the optimization process. Subsequently, the system uses a meta-reinforcement learning framework to process the aggregated feedback information in real time, dynamically adjusting the reasoning path and model parameters to adapt to the diverse needs of users. Based on user feedback, the system further updates the model parameters through online supervised learning to continuously improve the stability and long-term performance of the reasoning path. Through this closed-loop optimization mechanism, the system can accurately meet user needs and achieve intelligent knowledge exploration and text generation services.

[0169] The multi-person feedback aggregation algorithm based on game theory is described as follows:

[0170] To address the multi-user feedback conflict problem, a Shapley value feedback aggregation mechanism is proposed to ensure the fairness and stability of the optimization process:

[0171] Feedback contribution quantification:

[0172]

[0173] where φ u (v) represents the feedback contribution of user u, v(S) represents the feedback value of user subset S, and H represents the set of all users participating in feedback. represents the factorial.

[0174] When the query contains multiple modalities (such as text, image, audio, etc.), the system fuses various modal information by introducing a joint embedding space. For example, assuming T represents the text modality and I represents the image modality, the system can express both modalities in the following joint embedding e c :

[0175] e c = f(T, I) = α · f T (T) + (1 - α) · f I (I)

[0176] where f T (T) and f I (I) are the embedding representations of text and image, respectively, and α ∈ [0.1] is the weighting coefficient, used to control the degree of information fusion between text and image. This fusion method ensures that different modal information can jointly influence path selection in the reasoning path, thereby improving the flexibility and accuracy of the reasoning chain.

[0177] Vector representation: Word embedding and semantic embedding techniques not only convert entities into vectors but also capture semantic relationships between entities. For example, the relationship between "company" and "legal person" will have a high similarity in the vector space, while the relationship between "company" and "fine" will have a low similarity. The system measures the similarity between these vectors to determine the semantic relationship between entities, providing a basis for subsequent reasoning and thought chain generation.

[0178] The system adjusts the vector representation based on the input query and context. When processing user queries, the model can dynamically update the vector based on user feedback to optimize semantic understanding. For example, the user query "scope of company responsibility", the system will understand the specific meaning of "company" and "responsibility" in different legal fields (such as civil law or criminal law) through the large model, and adjust the vector representation to generate a more appropriate thought path.

[0179] When facing new tasks, the system quickly adjusts the reasoning path through transfer learning and meta-learning techniques. Through MAML in meta-learning, the reasoning path can quickly adapt to new query requirements from a small amount of task data:

[0180]

[0181] θ * represents the optimization of model parameters θ through meta-learning, so that the expected loss is minimized on the task distribution d(T). d(T) is the task distribution, representing the probability distribution of all possible tasks. ET~d(T) represents the loss L T of all tasks T.

[0182] where θ is the model parameter, θ' is the model parameter after task adaptation, and L T is the loss function of task T. By training an initial parameter θ, the model can quickly adjust to the optimal solution when encountering new tasks through a small amount of data.

[0183] Referring to Figure 2 , the entire technical process from multi-modal input to final output result is described, and how to optimize this process through feedback:

[0184] 1. Multi-modal input: Receive input data from different modalities, such as text, images and voice.

[0185] 2. Information structuring: The input information is structured to facilitate subsequent processing. This step ensures the organization and format of the information, making it suitable for machine processing.

[0186] 3. Task identification: Identify the specific task requirements of the user to provide direction for subsequent reasoning and processing.

[0187] 4. Vectorization: Information is converted into vector form to facilitate semantic processing. This step converts text, image, and speech data into numerical form that can be processed by machine learning models.

[0188] 5. Semantic Conflict Detection: Detect and handle semantic conflicts in input information. This step ensures consistency and accuracy of information.

[0189] 6. Dynamic Inference Path Generation: Generate dynamic inference paths based on task requirements. This step allows the system to dynamically adjust its inference strategy based on specific tasks.

[0190] 7. Dynamic Weight Assignment: Assign weights to individual nodes in the inference path to optimize the inference process. This step improves the efficiency and accuracy of inference by adjusting the importance of different inference steps.

[0191] 8. Meta Reinforcement Learning Optimization: Use meta reinforcement learning techniques to optimize the inference path, improving the adaptability and efficiency of the system. This step continuously improves the inference process by learning optimal strategies.

[0192] 9. Multi-level Inference: Perform multi-level inference to arrive at the final result. This step improves the depth and breadth of inference by analyzing and synthesizing information at different levels.

[0193] 10. Inference Process Visualization: Visualize the inference process for user understanding. This step improves the interpretability of results by graphically displaying the inference process.

[0194] 11. Output Results: Output the inference results to the user. This step ensures that the results are clear and easy to understand.

[0195] 12. Counterfactual Path Analysis: Analyze possible counterfactual paths to optimize the inference process. This step improves the flexibility and adaptability of the system by exploring different inference paths.

[0196] 13. User Interaction Feedback: Collect user feedback, including explicit and implicit signals. This step provides a basis for system optimization by collecting user feedback on results.

[0197] 14. Multi-party Feedback Processing: Process user feedback from different parties to further optimize the system. This step improves the comprehensiveness and accuracy of the system by integrating feedback from different sources.

[0198] 15. Structured Decision Recommendations: Provide structured decision recommendations based on inference results. This step improves the practicality of results by converting inference results into specific action recommendations.

[0199] 16. Reasoning process backtracking: allows users and the system to backtrack the reasoning process for understanding and improvement. This step improves the transparency and interpretability of the system by providing detailed records of the reasoning process.

[0200] The present application deeply integrates the reasoning path generation mechanism with dynamic weight distribution, meta-reinforcement learning, reasoning path optimization framework based on meta-reinforcement learning, causal reasoning, multi-person feedback aggregation algorithm based on game theory, and explainability enhancement technology, significantly improving the performance of intelligent query and knowledge reasoning system, and providing reliable technical support for knowledge service in complex scenarios.

[0201] 1) Reasoning path generation mechanism with dynamic weight distribution:

[0202] Through dynamic weight distribution, the present application can intelligently allocate the attention of different reasoning paths, prioritize key nodes, and ensure the logic and efficiency of the reasoning process. This solves the problem of missing key points caused by fixed weights in traditional methods.

[0203] 2) Meta-reinforcement learning and reasoning path optimization framework:

[0204] The meta-reinforcement learning framework allows the system to automatically adjust the reasoning strategy during operation, adapting to dynamically changing input scenarios. This mechanism can quickly optimize the reasoning path in real-time environments, improving accuracy and response speed.

[0205] 3) Causal reasoning and multi-dimensional correlation analysis:

[0206] Based on causal reasoning technology, the system can identify the true causal relationship between variables, avoiding false reasoning conclusions due to correlation. This is particularly important for handling multiple causal relationships in complex texts.

[0207] 4) Multi-person feedback aggregation algorithm based on game theory:

[0208] To meet the diverse needs of different users, the system uses game theory to balance multi-party feedback and find the optimal reasoning result. This method can coordinate the views of different stakeholders and improve the universality of the final conclusion.

[0209] 5) Explainability enhancement technology:

[0210] Through explainability enhancement technology, not only the final conclusion is output, but also the reasoning process and basis are detailed, improving the transparency and credibility of the results.

[0211] Based on the same inventive concept, the embodiments of the present application also provide a knowledge exploration system based on generative thinking chain and feedback mechanism, as shown in Figure 3 , comprising:

[0212] a data input module configured to receive a multimodal query input by a user, perform intent analysis to identify a core requirement, a writing topic, or a problem focus;

[0213] a semantic processing module configured to preprocess the multimodal query and extract key information, and decompose a complex problem into multiple sub-problems or sub-tasks;

[0214] a reasoning engine module configured to dynamically construct a hierarchical reasoning path based on a generative thinking chain according to the multiple sub-problems or sub-tasks, and label a logical basis for each step of the reasoning path;

[0215] a dynamic reasoning path module configured to optimize a priority of the reasoning path through dynamic weight allocation, and dynamically adjust a reasoning strategy online in combination with a meta-reinforcement learning framework to optimize the reasoning path;

[0216] a causal reasoning analysis module configured to, based on the optimized reasoning path, verify the reasoning path using a causal reasoning analysis tool, and generate an alternative path to correct an error node;

[0217] a result output and feedback optimization module configured to output a structured text and a visual reasoning process, and optimize model parameters in real time based on a multi-person feedback aggregation according to game theory.

[0218] As shown in Figure 3 , the system architecture depicts a complete process from data input to user feedback, as well as the interaction and data flow among the various modules.

[0219] Among them, the data input module is the starting point of user interaction with the system, responsible for receiving user query requests, which can be multimodal data in the form of text, image or voice. And SchemaNER entity recognition and structuring; After data input, the system uses SchemaNER technology to identify and structure the entities in the input data, and performs intent analysis to identify the core requirement, writing topic or problem focus.

[0220] The semantic processing module can perform dynamic vectorization processing on the input information, and is responsible for converting the user's query into a machine-understandable semantic vector. Through multi-modal joint embedding space technology, the system can process the fusion of text, image and voice, and realize the unified representation of information. It can preprocess the multimodal query and extract key information, and decompose complex problems into multiple sub-problems or sub-tasks.

[0221] The reasoning engine module is configured to generate a thinking chain; This module is the core of the system, responsible for generating a reasoning path. It can dynamically construct a hierarchical reasoning path based on a generative thinking chain according to multiple sub-problems or sub-tasks, and label a logical basis for each step of the reasoning path.

[0222] A dynamic reasoning path module dynamically generates reasoning paths based on task requirements through dynamic reasoning path tree generation techniques and assigns node weights. In combination with a meta-reinforcement learning framework, the reasoning strategy is adjusted online and dynamically to optimize the reasoning path. In addition, the system also expands the breadth and depth of reasoning through intent development and dispersion thinking techniques to derive the reasoning result.

[0223] A causal reasoning analysis module, based on the optimization of the reasoning path, uses causal reasoning analysis tools to verify the reasoning path and generate counterfactual paths to correct error nodes.

[0224] A result output and feedback optimization module converts the reasoning result into a structured explanation and presents it in a user-friendly manner, making the result easy to understand and operate. User feedback is collected (explicit / implicit signals) to collect feedback information from users. These feedbacks can be explicit (such as direct evaluation or suggestion) or implicit (such as click behavior or usage pattern). The feedback optimization part includes a meta-reinforcement learning optimizer (Meta-RL policy update), a supervised learning corrector (annotated data fine-tuning), and a causal counterfactual analysis (error path backtracking). These techniques are used to optimize the reasoning path and model parameters in real time, improving the adaptability and accuracy of the system.

[0225]

[0226] The knowledge exploration system based on generative thinking chain and feedback mechanism provided by the application is designed for complex long text scenarios and is suitable for formal and rigorous long text processing such as legal regulations, academic papers, technical specifications / standard documents, government reports, contract files, medical guidelines, financial reports, and user manuals. The following four application examples illustrate the scheme of the application:

[0227] Example 1: Intelligent search and retrieval system for regulations

[0228] Traditional regulation information retrieval systems usually rely on simple keyword matching, which results in low relevance and accuracy of query results, especially when facing complex legal issues, it is difficult to filter out the truly required clauses.

[0229] When the application is used to search or retrieve regulations, the system uses a dynamic weight allocation reasoning path generation mechanism combined with semantic understanding and context analysis capabilities of regulations to intelligently generate multi-level reasoning paths based on user input query requirements. The system can expand the reasoning path from multiple dimensions, such as the scope of application of legal clauses, effective conditions, and exceptional circumstances, to help users quickly locate key legal clauses and related cases. Through dynamic weight allocation, the system can prioritize important clauses and reduce irrelevant information interference to achieve more accurate retrieval results.

[0230] ​In addition, the system can also utilize meta-reinforcement learning techniques to optimize the reasoning path based on user feedback. For example, in cases where the user labels certain results as "irrelevant" or "relevant", the system adjusts the weight distribution strategy in real-time, gradually optimizing the accuracy of search results.

[0231] Embodiment 2: Legal Q&A System

[0232] In the legal Q&A scenario, the questions raised by users often involve complex legal backgrounds, which may require the analysis of multiple laws, regulations, and judicial interpretations. Traditional systems are difficult to associate multiple regulations in the same query, and cannot deeply explore the logical relationship between laws, resulting in incomplete answers.

[0233] When the application is used as a legal Q&A system, the system can automatically generate a dynamic reasoning path based on the user's query through dynamic reasoning path generation technology, and combine regulations and rights and responsibilities to form a multi-level legal analysis. For example, when dealing with "enterprise compliance review" related issues, the system can not only quote relevant provisions of the "Company Law", but also combine relevant provisions of the "Administrative Punishment Law" and the "Civil Procedure Law" to build a complete legal reasoning chain.

[0234] In addition, the system also supports a multi-round feedback optimization mechanism. Users can rate the answers or make additional demands, and the system will dynamically adjust the reasoning path based on feedback to gradually improve the quality and individualization of the answers. This method is particularly suitable for handling complex legal matters such as intellectual property disputes and contract performance disputes.

[0235] Embodiment 3: Legal Document Retrieval and Analysis System

[0236] In the face of a large amount of legal documents (such as cases, regulations interpretation, lawyer's opinion, etc.), traditional systems usually lack deep reasoning ability and semantic understanding ability, making it difficult to quickly locate the most relevant documents.

[0237] When using the system of the application for legal document retrieval and analysis, intelligent query path generation and causal reasoning technology can be used to identify the logical relationship between documents, such as the citation relationship of legal provisions, the correspondence between cases and legal provisions, etc., helping users quickly obtain the required information. For example, when searching for applicable cases of a certain legal principle, the system can combine causal reasoning to filter out relevant real cases and sort them by importance.

[0238] Through a dynamic feedback mechanism, the system can also make intelligent recommendations based on user preferences and behavior habits (such as frequently consulted regulatory areas, commonly used legal terms, etc.), further improving search efficiency and reducing the interference of invalid information.

[0239] Embodiment 4: Legal Consultation and Intelligent Customer Service

[0240] In legal consultation services, users may need personalized legal advice for specific issues. Traditional legal consultation systems usually only provide static knowledge base content and cannot dynamically adjust according to the specific circumstances of the user.

[0241] When using the system for legal consultation, the system can generate personalized legal advice based on the user's query history and behavior data by combining a feedback mechanism and a dynamic optimization framework. For example, when a user asks "how to handle a labor arbitration case", the system will not only quote relevant provisions of the Labor Law, but also combine local judicial practice and similar cases to provide more specific advice.

[0242] In addition, the system supports a multi-round interaction mode, which can gradually refine the user's query requirements in the conversation and ultimately generate a complete legal analysis report. Users can also score the reasoning results at each step, and the system will adjust the reasoning strategy for the next step based on the feedback to continuously improve service quality.

[0243] The present invention introduces a feedback mechanism based on meta-reinforcement learning to realize real-time reasoning path optimization. Through user feedback, the system can dynamically adjust the weight distribution of the reasoning path to optimize the accuracy of the query results. For example, the system can update the weight parameters in real time by user scoring or behavior data (such as the number of clicks to view) to gradually improve the relevance of the search results. To address the conflict problem of multiple user feedback, the system uses a Shapley value feedback aggregation mechanism to ensure the fairness and stability of the optimization process. By quantifying the feedback contribution of each user, the system can exclude the interference of low-quality feedback and optimize the selection strategy of the reasoning path.

[0244] In addition, the system also supports feedback-driven online supervised learning, which uses user feedback to continuously update model parameters to improve the stability and long-term performance of the reasoning path. For example, the system can generate adversarial samples after each user feedback to simulate the reasoning path in extreme situations and correct potential reasoning biases.

[0245] The various embodiments in the specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be mutually referred to. For the device disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method part.

[0246] The foregoing description of the disclosed embodiments enables a person skilled in the art to make or use the application. Modifications of these embodiments will occur to persons of skill in the art, and that the appended claims are intended to cover all such modifications that do not depart from the true spirit and scope of the application. Therefore, the application is not limited to the embodiments shown but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A knowledge exploration method based on generative thinking chains and feedback mechanisms, characterized in that, Includes the following steps: S10. Receive multimodal queries from user input, perform intent analysis to identify core needs, writing topics, or key issues; S20. Preprocess the multimodal query and extract key information, decomposing the complex problem into multiple sub-problems or sub-tasks; S30. Based on the multiple sub-problems or sub-tasks, dynamically construct a hierarchical reasoning path based on a generative thinking chain, and label the logical basis for each step of the reasoning path. S40. Optimize the priority of inference paths through dynamic weight allocation, and combine the meta-reinforcement learning framework to dynamically adjust the inference strategy online to optimize the inference path; S50. Based on the optimization of the reasoning path, use causal reasoning analysis tools to verify the reasoning path and generate counterfactual paths to correct erroneous nodes. S60 outputs structured text and visualizes the reasoning process, and optimizes model parameters in real time based on multi-person feedback aggregation in game theory; In step S40, optimizing the inference path priority through dynamic weight allocation includes: Based on the task context, semantic similarity, and node importance, the weight of each node in the inference path is dynamically calculated using the following formula: In the formula, w i Represents the reasoning node n i The weight; q represents the current query task; n i sim(q,n) represents the i-th inference node; j represents the total number of inference nodes in the current path; i ) represents the current task q and the inference node n i The semantic similarity; τ represents the temperature coefficient, used to control the intensity of path exploration; Based on the multiple nodes decomposed into the inference path, each node corresponds to a subtask, and the optimal path selection is achieved by maximizing the total reward value through a reward model; the formula for maximizing the total reward value is as follows: Where T represents the query task, R i (T) represents the reward value of a node, k represents the number of nodes in the current path, and p represents the dynamically adjusted normalized weight coefficient, p∈[0,1].

2. The knowledge exploration method based on generative thinking chain and feedback mechanism according to claim 1, characterized in that, Step S40, which optimizes the inference path priority through dynamic weight allocation, further includes: By fusing multimodal input information, semantic alignment of text and images is achieved through an alignment loss function, which is defined as follows: Among them, II ab ∈{0,1} represents the modal association indicator function, when text T a With Image I b The value is 1 if relevant, otherwise 0; f T (·) and f I (·) represent the text encoder and image encoder, respectively, outputting the normalized embedding vector; I k Let represent the k-th sample in the image modality, k≠b; cos(·,·) represents the cosine similarity, τ represents the temperature coefficient; N represents the number of text samples, and M represents the number of image samples.

3. The knowledge exploration method based on generative thinking chain and feedback mechanism according to claim 1, characterized in that, In step S40, the meta-reinforcement learning framework dynamically adjusts inference measurements online, including: Offline meta-pre-training: The model is pre-trained on a cross-domain task set using the MAML framework. Online fine-tuning generates adversarial examples based on real-time user feedback, dynamically updating model parameters. A composite reward mechanism that integrates interpretability metrics balances inference accuracy, efficiency, and interpretability through a multi-objective reward function design.

4. The knowledge exploration method based on generative thinking chain and feedback mechanism according to claim 1, characterized in that, In step S50, the reasoning path is verified using a causal reasoning analysis tool, including: When the reasoning result deviates, a counterfactual query is generated to compare the actual path with the counterfactual path to locate the erroneous node and trigger correction. By embedding causal probability edges into knowledge to define the causal strength between nodes, and calculating logical consistency scores, task relevance scores, case support scores, and interpretability scores, the credibility and interpretability of reasoning are enhanced.

5. A knowledge exploration method based on generative thinking chain and feedback mechanism according to claim 4, characterized in that, In step S50, verifying the reasoning path using a causal reasoning analysis tool further includes: At each step of the inference path generation, the inference chain is optimized using a reward model, with the following reward mechanism: R=w1·Accuracy+w2·Efficiency-w3·Resources w1 represents the weighting coefficient for accuracy; Accuracy represents the accuracy score of the inference chain; w2 represents the weighting coefficient for efficiency; Efficiency represents the efficiency score of the inference chain; w3 represents the weighting coefficient for resource consumption; Resources represents the resource consumption score of the inference chain.

6. The knowledge exploration method based on generative thinking chain and feedback mechanism according to claim 1, characterized in that, In step S60, the real-time optimization of model parameters based on game theory-based multi-person feedback aggregation includes: The Shapley value algorithm, based on game theory, quantifies the contribution of user feedback, filters low-quality or malicious feedback, and coordinates the opinions of multiple users to optimize the inference path. The meta-reinforcement learning framework is used to process the aggregated feedback information in real time and dynamically adjust the inference path and model parameters.

7. A knowledge exploration system based on generative thinking chains and feedback mechanisms, characterized in that, The system, employing the knowledge exploration method as described in any one of claims 1-6, comprises: The data input module is used to receive multimodal queries from users, perform intent analysis to identify core needs, writing topics, or key issues; The semantic processing module is used to preprocess the multimodal query and extract key information, decomposing complex problems into multiple sub-problems or sub-tasks; The reasoning engine module is used to dynamically construct a hierarchical reasoning path based on the multiple sub-problems or sub-tasks, and to annotate the logical basis for each step of the reasoning path. The dynamic inference path module is used to optimize the priority of inference paths through dynamic weight allocation, and combines the meta-reinforcement learning framework to dynamically adjust the inference strategy online to optimize the inference path; The causal reasoning analysis module is used to verify the reasoning path using causal reasoning analysis tools, and generate counterfactual paths to correct erroneous nodes, based on the optimization of the reasoning path. The results output and feedback optimization module is used to output structured text and visualize the reasoning process, and optimize the model parameters in real time based on game theory-based multi-person feedback aggregation.

Images