Meta-cognition space division-based large language model knowledge enhancement method and system
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HARBIN INST OF TECH
- Filing Date
- 2026-04-22
- Publication Date
- 2026-08-07
AI Technical Summary
[0004]本发明为解决现有技术中因忽视大语言模型内在认知状态而导致的干预盲目性、知识增强效率低下，以及缺乏有效的主客观一致性校准手段的问题，进而提出基于元认知空间划分的大语言模型知识增强方法及系统
1.本发明通过引入一个创新的、基于模型内部信号（如负对数似然）的认知空间划分机制，准确且低成本地量化模型对特定知识点的掌握程度，解决了模型“主观自信与客观准确性脱节”的深层矛盾；通过实施差异化的“诊断-扩展-校准”迭代机制，巧妙地平衡了外部知识的引入与内部认知的校准，克服了传统方法干预无序、噪声引入过多的局限性，最终显著提升大语言模型在知识密集型任务中的推理性能与决策可靠性，实现“知其然且知其所以然”的增强目标。
Smart Images
Figure CN122529062A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a method and system for knowledge enhancement of large language models based on metacognitive space partitioning, belonging to the fields of natural language processing and artificial intelligence technology. Background Technology
[0002] Knowledge augmentation, a key paradigm for improving the performance of LLMs in knowledge-intensive tasks, compensates for the model's shortcomings by introducing external information or adjusting internal parameters. However, existing technologies for improving model performance often rely on an overly idealistic premise: simply equating the model's output prediction performance with its internal knowledge reserves.
[0003] Currently, the most similar implementations to this invention primarily focus on triggering a knowledge supplementation mechanism by evaluating the correctness of prediction results. This involves blindly supplementing the model's incorrect knowledge points through methods such as retrieval enhancement or supervised fine-tuning. The drawback of this approach is its singular cognitive evaluation dimension, neglecting the complex "knowledge-confidence" gap within the model, making it highly susceptible to cognitive dissonance when processing tasks. Specifically, this approach cannot distinguish between accidental errors based on prior knowledge and hallucinations caused by a knowledge vacuum. This intervention mechanism lacks specificity and cannot differentiate interventions based on the model's different internal cognitive states (such as mastery, confusion, or lack of knowledge). Furthermore, because such approaches fail to establish effective alignment between subjective and objective factors, the model may still maintain overconfident erroneous cognitions or exhibit uncertain correct responses after receiving knowledge supplementation, severely limiting the reliability and robustness of knowledge enhancement in complex reasoning scenarios. Summary of the Invention
[0004] To address the problems of blind intervention, low efficiency of knowledge enhancement, and lack of effective subjective-objective consistency calibration methods in existing technologies due to neglecting the inherent cognitive state of large language models, this invention proposes a knowledge enhancement method and system for large language models based on metacognitive space partitioning.
[0005] The technical solution adopted by this invention to solve the above problems is: the knowledge enhancement method for large language models based on metacognitive space partitioning proposed in this invention includes: Step 1: Generate multiple response paths based on the input query, and obtain the internal uncertainty score for each response path; Step 2: Based on the internal uncertainty score and the predictive consistency of the response path, the knowledge space is mapped into three cognitive regions: the mastery region, the confusion region, and the missing region. Step 3: Perform differentiated knowledge expansion operations on the large language model based on the category of cognitive regions; Step 4: During the training process after the knowledge expansion operation, the entropy value of the output probability distribution is dynamically adjusted according to the correctness of the reasoning results to calibrate the cognitive confidence of the model, and the closed-loop iterative execution is carried out to complete the knowledge enhancement of the large language model.
[0006] Furthermore, step 1 specifically includes: Calculate the average negative log-likelihood of the response path based on the conditional generation probability of each tag when generating the response path using the large language model. The average negative log-likelihood value As an internal uncertainty score, the internal uncertainty score is obtained by calculating only the underlying probability distribution of the large language model itself.
[0007] Furthermore, the average negative log-likelihood value The calculation formula is: (1); In formula (1), The model represents the generation of the first query based on the given context and query conditions. i The probability of each label. L Indicates the length of the response path.
[0008] Furthermore, step 2 specifically includes: When the internal uncertainty score is below the first threshold and the prediction results of multiple response paths are consistent, the corresponding knowledge area is divided into the mastery area. When the internal uncertainty score is between the first threshold and the second threshold, or when the internal uncertainty score is lower than the first threshold but the prediction results of multiple response paths are inconsistent, the corresponding knowledge area is divided into a confusion area. When the internal uncertainty score is higher than the second threshold, the corresponding knowledge area is divided into a missing area; where the first threshold is less than the second threshold.
[0009] Furthermore, step 3 specifically includes: When the cognitive region is a missing region, authoritative facts are obtained through automated retrieval and transformed into instantiated question-and-answer training pairs to fill the knowledge gaps; When the cognitive region is a region of confusion, a chain of reasoning supported by explicit evidence is introduced to eliminate semantic ambiguity; When the cognitive region is the mastery region, training data is expanded by introducing background questions to broaden the knowledge boundary.
[0010] Furthermore, step 4 dynamically adjusts the entropy value of the output probability distribution based on the correctness of the reasoning result, including: When the reasoning result is correct, reduce the entropy value of the output probability distribution to enhance the model's answer confidence. When the reasoning result is incorrect, the entropy value of the output probability distribution is increased to induce cognitive hesitation in the model.
[0011] Furthermore, in step 4, the closed-loop iterative execution includes three stages: detection, intervention, and calibration. The detection stage includes generating a response path and calculating the internal uncertainty score; the intervention stage includes dividing the cognitive region and performing differentiated knowledge expansion operations; and the calibration stage includes dynamically adjusting the entropy value of the output probability distribution. The closed-loop iteration continues until the preset termination condition is met.
[0012] Furthermore, this invention also proposes a knowledge enhancement system for large language models based on metacognitive space partitioning, including: The metacognitive detection module is used to generate multiple response paths based on the input query and obtain the internal uncertainty score of each response path; The cognitive space partitioning module is used to map the knowledge space into three cognitive regions: the mastery region, the confusion region, and the missing region, based on the internal uncertainty score and the predictive consistency of the response path. The targeted knowledge expansion module is used to perform differentiated knowledge expansion operations on the large language model based on the category of the cognitive region; The cognitive consistency calibration module is used to dynamically adjust the entropy value of the output probability distribution based on the correctness of the reasoning results during the training process after knowledge expansion operations, in order to calibrate the cognitive confidence of the model.
[0013] The beneficial effects of this invention are: 1. This invention introduces an innovative cognitive space partitioning mechanism based on internal model signals (such as negative log-likelihood) to accurately and cost-effectively quantify the model's mastery of specific knowledge points, thus resolving the deep-seated contradiction of "disconnect between subjective confidence and objective accuracy" in the model. By implementing a differentiated "diagnosis-expansion-calibration" iterative mechanism, it cleverly balances the introduction of external knowledge with the calibration of internal cognition, overcoming the limitations of traditional methods such as disordered intervention and excessive noise introduction. Ultimately, it significantly improves the reasoning performance and decision reliability of large language models in knowledge-intensive tasks, achieving the enhancement goal of "knowing not only what but also why".
[0014] 2. This invention constructs a task-agnostic, fully automated metacognitive knowledge enhancement framework. Through a three-stage closed-loop iteration of "detection-intervention-calibration", this framework can accurately identify and repair knowledge defects within the model without manual annotation.
[0015] 3. This invention proposes a novel metacognitive diagnostic method based on the structural decay law. This method establishes an exponential mapping relationship between subjective uncertainty and objective accuracy by quantifying the negative log-likelihood (NLL) signal inside the model, thereby achieving efficient division of the mastery area, confusion area and missing area, and its computational cost is far lower than that of evaluation methods based on external models or complex sampling.
[0016] 4. This invention establishes a differentiated intervention and entropy control calibration mechanism based on cognitive regions. This mechanism expands knowledge in a targeted manner (such as instantiation completion and disambiguation reasoning) to different cognitive regions, and synchronizes confidence and accuracy by dynamically adjusting the entropy value of the predicted probability distribution during the training phase. This significantly improves the model's knowledge reserves while effectively solving the balance problem between "illusion caused by overconfidence" and "instability caused by cognitive dissonance" in existing technologies, ensuring the reliability of the model output from the underlying probabilistic logic. Attached Figure Description
[0017] Figure 1 A flowchart illustrating a knowledge enhancement method for large language models based on metacognitive space partitioning; Figure 2 This is a framework diagram for a large language model knowledge enhancement system based on metacognitive space partitioning. Detailed Implementation
[0018] like Figure 1 As shown, the steps of the large language model knowledge enhancement method based on metacognitive space partitioning described in this embodiment include: S1: Generate multiple response paths based on the input query and obtain the internal uncertainty score for each response path; Before detailing the process, it's necessary to introduce the core metric of this invention: a metacognitive metric based on the model's internal probability signals. This metric uses negative log-likelihood (NLL) to quantify the model's internal uncertainty in generating a specific response. This invention discovers and utilizes the "structural decay law," which states that there is a significant exponential negative correlation between the model's objective accuracy and its subjective uncertainty. Specifically, for a given query... q The generated length is L response path r Its internal uncertainty score u The calculation formula is: (1); In formula (1), The model represents the generation of the first query based on the given context and query conditions. i The probability of each label.L This indicates the length of the response path. Using this metric, the system can directly utilize the underlying probability distribution of the large language model itself for calculation, without introducing an additional external evaluation model. This avoids errors caused by embedding space mismatch and additional computational overhead, achieving both high diagnostic efficiency and accuracy.
[0019] S2: Divide the knowledge space mapping region according to the internal uncertainty score; The overall method of this invention continuously optimizes the knowledge state of the model through an iterative cycle of detection and intervention. In the cognitive space diagnosis step, the system performs multi-path random sampling on the input query and automatically maps the knowledge space into three heterogeneous regions—the mastery region, the confusion region, and the missing region—based on metrics such as prediction consistency, uncertainty, and accuracy, combined with preset thresholds. The mastery region corresponds to the knowledge that the model has robustly mastered, the confusion region corresponds to knowledge with cognitive dissonance or logical fragility, and the missing region identifies knowledge vacuums within the model.
[0020] S3: Perform differentiated knowledge expansion operations on the large language model based on the category of the cognitive region; While there are alternative solutions that can achieve similar objectives to the technical solution of this invention, they all suffer from significant shortcomings in terms of performance accuracy or implementation cost. In the cognitive measurement stage, although methods such as semantic consistency, hidden layer feature clustering, or introducing an external model as a judge (LLM-as-Judge) can be used to replace the negative log-likelihood-based quantization scheme of this invention, these alternative paths are generally weaker in detection accuracy than this invention, and some schemes have enormous computational overhead, making it difficult to meet the requirements of efficient enhancement.
[0021] In the knowledge expansion and calibration steps, this invention employs a differentiated intervention mechanism to balance the depth of knowledge with the consistency of cognition. For missing areas, the system obtains authoritative facts through an automated retrieval mechanism and transforms them into instantiated QA training pairs to fill the underlying knowledge gaps. For confused areas, it introduces reasoning chains supported by explicit evidence to eliminate semantic ambiguity and solidify logical connections. For mastered areas, it expands the knowledge boundaries of mastered knowledge by enriching training data with derived question backgrounds. Simultaneously, this scheme introduces a unique cognitive-driven alignment mechanism: during reinforcement learning optimization, the system dynamically adjusts the entropy value of the output probability distribution based on the correctness of the reasoning results. For correct paths, it strengthens the model's answer certainty by suppressing entropy; for incorrect paths, it induces "cognitive hesitation" in the model by increasing entropy. This design avoids the illusion of overconfidence caused by blindly adding knowledge, ensuring that the model can simultaneously calibrate its subjective and objective cognitive boundaries while acquiring new knowledge.
[0022] S4: Perform closed-loop iterative execution to complete the knowledge enhancement of the large language model.
[0023] While intervention and calibration methods can attempt to improve performance through full fine-tuning, indiscriminate retrieval enhancement, or simple output filtering, these approaches often fail to simultaneously ensure accurate knowledge filling and deep calibration with consistency between subjective and objective perspectives. In contrast, this invention, through spatial partitioning and differentiated intervention combined with the closed-loop logic of entropy control calibration, possesses comprehensive advantages unmatched by existing alternatives in achieving high-quality, highly consistent knowledge enhancement. Even if some components are replaced with the aforementioned prior art in specific implementations, as long as they fall within the metacognitive architecture logic of this invention ("diagnosis-intervention-calibration"), they should still fall within the protection scope of this invention.
[0024] The closed-loop iterative execution comprises three stages: detection, intervention, and calibration. The detection stage involves generating response paths and calculating internal uncertainty scores; the intervention stage involves dividing cognitive regions and performing differentiated knowledge expansion operations; and the calibration stage involves dynamically adjusting the entropy value of the output probability distribution. The closed-loop iteration continues until a preset termination condition is met. Through continuous iteration of "diagnosis-intervention-calibration," the model's internal knowledge reserves and cognitive confidence are synergistically optimized, forming a metacognitive system that is both knowledgeable and prudent, significantly improving the reliability of large language models in complex application scenarios.
[0025] In addition, such as Figure 2 As shown, this embodiment also proposes a large language model knowledge enhancement system based on metacognitive space partitioning, which includes a metacognitive detection module, a cognitive space partitioning module, a targeted knowledge expansion module, and a cognitive consistency calibration module connected in sequence.
[0026] The metacognitive detection module generates multiple response paths based on the input query and obtains the internal uncertainty score for each response path. The cognitive space partitioning module maps the knowledge space into three cognitive regions: mastery region, confusion region, and missing region, based on the internal uncertainty score and the prediction consistency of the response paths. The targeted knowledge expansion module performs differentiated knowledge expansion operations on the large language model according to the category of the cognitive region. The cognitive consistency calibration module dynamically adjusts the entropy value of the output probability distribution based on the correctness of the reasoning results during the training process after the knowledge expansion operation, so as to calibrate the cognitive confidence of the model.
[0027] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent substitutions, and improvements made to the above embodiments without departing from the scope of the present invention, based on the technical essence of the present invention and within the spirit and principles of the present invention, shall still fall within the protection scope of the present invention.
Claims
1. A knowledge enhancement method for large language models based on metacognitive space partitioning, characterized in that, include: Step 1: Generate multiple response paths based on the input query, and obtain the internal uncertainty score for each response path; Step 2: Based on the internal uncertainty score and the predictive consistency of the response path, the knowledge space is mapped into three cognitive regions: the mastery region, the confusion region, and the missing region. Step 3: Perform differentiated knowledge expansion operations on the large language model based on the category of cognitive regions; Step 4: During the training process after the knowledge expansion operation, the entropy value of the output probability distribution is dynamically adjusted according to the correctness of the reasoning results to calibrate the cognitive confidence of the model, and the closed-loop iterative execution is carried out to complete the knowledge enhancement of the large language model.
2. The knowledge enhancement method for a large language model based on metacognitive space partitioning according to claim 1, characterized in that, Step 1 specifically includes: Calculate the average negative log-likelihood of the response path based on the conditional generation probability of each tag when generating the response path using the large language model. The average negative log-likelihood value As an internal uncertainty score, the internal uncertainty score is obtained by calculating only the underlying probability distribution of the large language model itself.
3. The knowledge enhancement method for a large language model based on metacognitive space partitioning according to claim 2, characterized in that, The average negative log-likelihood value The calculation formula is: （1）； In formula (1), The model represents the generation of the first query based on the given context and query conditions. i The probability of each label. L Indicates the length of the response path.
4. The knowledge enhancement method for a large language model based on metacognitive space partitioning according to claim 1, characterized in that, Step 2 specifically includes: When the internal uncertainty score is below the first threshold and the prediction results of multiple response paths are consistent, the corresponding knowledge area is divided into the mastery area. When the internal uncertainty score is between the first threshold and the second threshold, or when the internal uncertainty score is lower than the first threshold but the prediction results of multiple response paths are inconsistent, the corresponding knowledge area is divided into a confusion area. When the internal uncertainty score is higher than the second threshold, the corresponding knowledge area is divided into a missing area; wherein the first threshold is less than the second threshold.
5. The knowledge enhancement method for a large language model based on metacognitive space partitioning according to claim 1, characterized in that, Step 3 specifically includes: When the cognitive region is a missing region, authoritative facts are obtained through automated retrieval and transformed into instantiated question-and-answer training pairs to fill the knowledge gaps; When the cognitive region is a region of confusion, a chain of reasoning supported by explicit evidence is introduced to eliminate semantic ambiguity; When the cognitive region is the mastery region, training data is expanded by introducing background questions to broaden the knowledge boundary.
6. The knowledge enhancement method for a large language model based on metacognitive space partitioning according to claim 1, characterized in that, Step 4 involves dynamically adjusting the entropy value of the output probability distribution based on the correctness of the inference result, including: When the reasoning result is correct, the entropy value of the output probability distribution is reduced to enhance the model's answer confidence. When the reasoning result is incorrect, the entropy value of the output probability distribution is increased to induce cognitive hesitation in the model.
7. The knowledge enhancement method for a large language model based on metacognitive space partitioning according to claim 1, characterized in that, Step 4 involves a closed-loop iterative execution comprising three phases: detection, intervention, and calibration. The detection phase includes generating a response path and calculating the internal uncertainty score; the intervention phase includes dividing the cognitive region and performing differentiated knowledge expansion operations; and the calibration phase includes dynamically adjusting the entropy value of the output probability distribution. The closed-loop iteration continues until the preset termination condition is met.
8. A knowledge enhancement system for large language models based on metacognitive space partitioning, applied to the knowledge enhancement method for large language models based on metacognitive space partitioning as described in any one of claims 1-7, characterized in that, include: The metacognitive detection module is used to generate multiple response paths based on the input query and obtain the internal uncertainty score of each response path; The cognitive space partitioning module is used to map the knowledge space into three cognitive regions: the mastery region, the confusion region, and the missing region, based on the internal uncertainty score and the predictive consistency of the response path. The targeted knowledge expansion module is used to perform differentiated knowledge expansion operations on the large language model based on the category of the cognitive region; The cognitive consistency calibration module is used to dynamically adjust the entropy value of the output probability distribution based on the correctness of the reasoning results during the training process after knowledge expansion operations, in order to calibrate the cognitive confidence of the model.