A Method for Optimizing and Dynamically Calibrizing Evaluation Models Based on Meta-Evaluation Feedback
By combining the collaborative architecture of the main evaluation model and the meta-evaluation model with a dynamic knowledge graph and a reinforcement learning framework, the rigidity and insufficient adaptability of the traditional evaluation system are solved, and the dynamic adjustment and stable performance output of the evaluation system are realized.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- INSPUR QILU SOFTWARE IND
- Filing Date
- 2025-06-24
- Publication Date
- 2026-07-03
AI Technical Summary
Traditional model evaluation systems suffer from rigid evaluation criteria and insufficient adaptability, making it difficult to dynamically adjust when models are iterated and upgraded or knowledge bases are dynamically expanded. This leads to a systematic deviation between the evaluation results and the actual performance of the model.
We adopt a collaborative architecture of main evaluation model and meta-evaluation model, combined with dynamic knowledge graph and reinforcement learning framework, and achieve error correction, feature optimization and robustness improvement through four-dimensional feedback mechanism and generative adversarial network.
It enables dynamic adjustment of the evaluation system, ensuring that the evaluation strategy and model are optimized in sync, and maintaining stable performance output during continuous iteration.
Smart Images
Figure CN120806034B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of artificial intelligence model evaluation, and in particular to a method for optimizing and dynamically calibrating evaluation models based on meta-evaluation feedback. Background Technology
[0002] Traditional model evaluation systems have long suffered from the dual dilemma of rigid evaluation criteria and insufficient adaptability. In particular, when the evaluated model undergoes iterative upgrades or the relevant domain knowledge base is dynamically expanded, the evaluation model struggles to dynamically adjust its judgment criteria, leading to a systematic deviation between the evaluation results and the model's actual performance. Summary of the Invention
[0003] To address the aforementioned technical problems, this invention provides a method for optimizing and dynamically calibrating an evaluation model based on meta-evaluation feedback. By introducing a meta-evaluation model and a four-dimensional feedback mechanism, combined with the real-time domain knowledge update capability of a dynamic knowledge graph and the construction of a parameter optimization loop using reinforcement learning strategies, the performance degradation problem caused by the lag in the evaluation system is resolved.
[0004] This invention addresses the issue of performance degradation caused by the lag in the evaluation system when the evaluated model undergoes iterative upgrades or the relevant domain knowledge base is dynamically expanded. By adopting a dual-model collaborative architecture and a four-dimensional feedback mechanism, combined with the real-time update capability of the domain knowledge of the dynamic knowledge graph and the construction of parameter optimization loops using reinforcement learning strategies, it solves the problem of performance degradation caused by the lag in the evaluation system and provides a reliable dynamic evaluation framework for the continuous iteration of artificial intelligence models.
[0005] The technical solution of this invention is:
[0006] An evaluation model optimization and dynamic calibration method based on meta-evaluation feedback is proposed. Through the collaborative operation of a main evaluation model and a meta-evaluation model, combined with a dynamic knowledge graph and a reinforcement learning framework, an evaluation system with real-time feedback optimization capabilities is constructed. The main evaluation model is responsible for performing the original answer matching task, while the meta-evaluation model achieves error correction and logical optimization through analysis of variance, semantic alignment, and adversarial verification. The two models form a mutual verification mechanism through adversarial training, and dynamic perturbations are injected using a generative adversarial network to improve the system's robustness.
[0007] Furthermore,
[0008] Synergy between the main evaluation model and the meta-evaluation model
[0009] The main evaluation model performs the original answer matching task.
[0010] The main evaluation model, as the basic execution layer, is responsible for directly processing input data and generating preliminary results;
[0011] The meta-evaluation model performs a secondary verification of the main model's judgment.
[0012] The meta-evaluation model enhances the reliability of the main model through multi-dimensional validation, including:
[0013] Error correction: Identify the bias of the master model through analysis of variance and Kendall's coefficient of harmony method, and dynamically adjust the weights;
[0014] Feedback optimization: Step-level feedback guides the iterative inference path of the main model;
[0015] Improve system robustness through adversarial training
[0016] Adversarial training enhances a model's resistance to noise and anomalous inputs by simulating attack scenarios. This can be achieved through methods such as:
[0017] Dynamic perturbation generation: By using generative adversarial networks (GANs) or projection adversarial training, small perturbations are injected into the input data, forcing the main model to learn more stable feature representations;
[0018] Dual-model cross-validation mechanism: The main model and the meta-model act as adversaries to each other during training;
[0019] Robustness evaluation metrics: Combine statistical testing to quantify the performance degradation of the model in adversarial environments, and guide the adjustment of training strategies.
[0020] Furthermore,
[0021] It adopts a deep integration of a four-dimensional feedback mechanism and a dynamic knowledge graph.
[0022] in,
[0023] The four-dimensional feedback mechanism includes:
[0024] Feature enhancement mechanisms optimize model convergence efficiency by enhancing high-frequency feature weights and suppressing interference.
[0025] The gradient penalty mechanism uses a reverse correction strategy to correct erroneous decisions;
[0026] Feature reconstruction mechanisms are used to specifically repair confused regions in the coding layer;
[0027] The dynamic threshold mechanism achieves threshold safety calibration based on a statistical monitoring window.
[0028] The dynamic knowledge graph, through the integrity constraints of 12,000+ answer nodes, industry feature weighting, and timeliness labeling, forms a closed-loop linkage with the evaluation system to generate synthetic data and enhance the knowledge evolution and adaptability.
[0029] Furthermore,
[0030] The reinforcement learning framework employs the PPO algorithm with multimodal input to achieve real-time feedback between the evaluation strategy and the knowledge graph.
[0031] Furthermore,
[0032] Feature enhancement mechanism:
[0033] When the main model correctly matches the answer and the meta-model confirms that its confidence level is reasonable, Toucht performs a feature enhancement process: extracting key features of the correct answer from the semantic encoding layer of the main model, comparing them with the feature library of historical correct samples, and identifying high-frequency activated feature dimensions; adding gradient update weights to the neuron connections corresponding to high-frequency features to make them converge faster during training; and reducing the weight of non-core features through random masking technology to reduce the interference of redundant information on decision-making.
[0034] Gradient penalty mechanism:
[0035] When the main model incorrectly rejects the correct answer, and the meta-model confirms the error through adversarial examples, the penalty strength is calculated based on the confidence level of the main model when it incorrectly rejects the answer. The higher the penalty strength, the greater the parameter correction. During the backpropagation phase of the main model, the gradient direction is reversed on the neuron connections that caused the incorrect rejection. The maximum magnitude of a single gradient update is set to avoid damaging the model stability due to excessive penalty.
[0036] Feature reconstruction mechanism:
[0037] When the main model incorrectly accepts an incorrect answer, and the meta-model detects such errors through adversarial examples, it performs cluster analysis on the semantic features of the incorrectly accepted answer to locate the confusion region in the main model's encoding layer; based on the number of incorrect samples, it selects specific channels in the encoding layer for reconstruction in proportion; it only unfreezes the neurons in the selected channels for fine-tuning, while keeping other parameters frozen to ensure that the main structure of the knowledge base is not affected;
[0038] Dynamic threshold mechanism:
[0039] The main model's judgment threshold is dynamically calibrated based on the meta-model's adversarial verification results. N test samples are set as windows, and the confidence difference between the main model and the meta-model is statistically analyzed. If the difference persists, the main model threshold is adjusted by a fixed step size. A threshold fluctuation range is set to prevent extreme adjustments. At the same time, a delayed effect mechanism is introduced, which only takes effect when the trends of three consecutive monitoring windows are consistent.
[0040] Furthermore,
[0041] Dynamic knowledge graph
[0042] Coupling of Answer Completeness Graph and Dual-Model Collaborative Architecture
[0043] In the original answer matching stage of the main evaluation model, 12,000+ pre-set answer nodes in the knowledge graph are used as the semantic integrity benchmark library; the answers are mapped to vectors through graph embedding technology, and graph structure similarity constraints are introduced when calculating similarity with user answers; the meta-evaluation model learns the matching deviation pattern of the main model through adversarial training, and verifies whether the main model has missed key knowledge branches by combining the node coverage status of the integrity graph.
[0044] Dynamic adaptation of feature weight map and four-dimensional feedback mechanism
[0045] Based on the industry feature weight map, a domain-sensitive feature enhancement mechanism is designed to synchronously associate the feature weight map: if the model's evaluation result in a specific domain deviates from the map weight distribution, the gradient penalty coefficient of the domain features is automatically increased to force the model parameters to converge to the weight distribution labeled in the map.
[0046] Optimization of dynamic threshold mechanism driven by time-sensitive personality spectrum
[0047] The daily / weekly / monthly frequency markers of the time-sensitive personality spectrum are linked with a dynamic threshold mechanism to design a time decay function; when the main model outputs results, the confidence threshold is dynamically adjusted in conjunction with the time-sensitive markers.
[0048] The entire closed loop of graph evolution mechanism and model parameters
[0049] An evaluation-graph-training closed loop is constructed. Meanwhile, the feature reconstruction mechanism is based on the evolved graph and uses a graph neural network (GNN) to encode node weights and connections, generating synthetic data with enhanced domain features. This data is then injected into the adversarial training set to improve the model's adaptability to knowledge changes.
[0050] Furthermore,
[0051] Reinforcement learning framework
[0052] Deep integration of reward function and evaluation system
[0053] The reward mechanism achieves multi-level reward determination through deep integration with the adversarial verification process of the dual-model collaborative architecture; and ensures the accuracy of reward signals and the rationality of exploration directions through dual verification of adversarial verification and dynamic threshold linkage.
[0054] Cooperative Implementation of PPO Algorithm and Model Architecture
[0055] The policy network adopts a multimodal input design, concatenating the main model confidence, meta-model validation features, four-dimensional feedback indicators, and knowledge graph embedding vectors into a 512-dimensional feature. Decisions are made through a 3-layer MLP with residual connections, i.e., 512-256-128 nodes. A feature enhancement gating mechanism is introduced at the 256-node layer to dynamically adjust the information flow intensity based on the feature weights in the four-dimensional feedback. The exploration strategy adopts a dynamic ε-greedy algorithm, with an initial 15% exploration rate that is automatically adjusted by statistically analyzing the effective exploration rate through a sliding window.
[0056] The beneficial effects of this invention are
[0057] It effectively solves the problems of rigid standards and lagging performance in traditional evaluation systems, enables dynamic adjustment of evaluation weights, and allows the standard system to evolve autonomously as domain knowledge expands.
[0058] In addressing performance lag, the evaluation strategy and model upgrades are optimized in sync to ensure that the evaluation system maintains stable performance output during continuous iteration. Attached Figure Description
[0059] Figure 1 This is a schematic diagram of the workflow of the present invention. Detailed Implementation
[0060] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.
[0061] This invention proposes an evaluation model optimization system and dynamic calibration method based on meta-evaluation feedback, aiming to solve the problems of rigid standards, insufficient adaptability, and performance lag in traditional model evaluation systems. This technical solution constructs an evaluation system with real-time feedback optimization capabilities through the collaborative operation of a main evaluation model and a meta-evaluation model, combined with dynamic knowledge graphs and reinforcement learning frameworks. The main evaluation model is responsible for performing the original answer matching task, while the meta-evaluation model achieves error correction and logical optimization through variance analysis, semantic alignment, and adversarial verification. The two models form a mutual verification mechanism through adversarial training, and dynamic perturbations are injected using generative adversarial networks to improve system robustness.
[0062] The core innovation lies in the deep integration of a four-dimensional feedback mechanism and a dynamic knowledge graph: the feature enhancement mechanism optimizes model convergence efficiency through high-frequency feature weight enhancement and interference suppression; the gradient penalty mechanism corrects erroneous decisions using a reverse correction strategy; the feature reconstruction mechanism repairs confused regions in the encoding layer; and the dynamic threshold mechanism achieves threshold safety calibration based on a statistical monitoring window. The dynamic knowledge graph, through the integrity constraints of over 12,000 answer nodes, industry feature weight adaptation, and timeliness labeling, forms a closed-loop linkage with the evaluation system, generating synthetic data to enhance the knowledge's evolutionary adaptability. The reinforcement learning framework employs the PPO algorithm with multimodal input to achieve real-time mutual feedback between the evaluation strategy and the knowledge graph.
[0063] The specific technical solution of the present invention is as follows:
[0064] b) Dual-model collaborative architecture
[0065] ④ The main evaluation model performs the original answer matching task.
[0066] The main evaluation model, acting as the foundational execution layer, is responsible for directly processing the input data and generating preliminary results. For example, in a question-answering system, the main model filters the original answers through semantic vector matching or keyword retrieval. These models typically require high accuracy but may have limitations due to their single perspective.
[0067] ⑤ The meta-evaluation model performs a secondary verification of the main model's judgment.
[0068] The meta-evaluation model (Critique Model) enhances the reliability of the main model through multi-dimensional validation. This includes:
[0069] Error correction: Identify the bias of the master model through analysis of variance and Kendall's coefficient of harmony method, and dynamically adjust the weights;
[0070] Feedback optimization: Step-by-step feedback guides the iterative reasoning path of the main model, avoiding logical errors caused by the limitations of a single model.
[0071] ⑥ Improve system robustness through adversarial training
[0072] Adversarial training enhances a model's resistance to noise and anomalous inputs by simulating attack scenarios. This can be achieved through methods such as:
[0073] Dynamic perturbation generation: By using generative adversarial networks (GANs) or projection adversarial training, small perturbations are injected into the input data, forcing the main model to learn more stable feature representations;
[0074] Dual-model mutual verification mechanism: The main model and the meta-model act as adversaries during training. For example, adversarial examples are generated by the main model, and the meta-model evaluates their rationality, forming a closed-loop optimization.
[0075] Robustness evaluation metrics: Combine statistical testing to quantify the performance degradation of the model in adversarial environments, and guide the adjustment of training strategies.
[0076] c) Four-dimensional feedback mechanism
[0077] ⑤ Feature enhancement mechanism:
[0078] When the main model correctly matches the answer and the meta-model confirms that its confidence level is reasonable, Toucht performs a feature enhancement process: extracting key features of the correct answer from the semantic encoding layer of the main model, comparing them with the feature library of historical correct samples, and identifying the feature dimensions of high-frequency activation; adding gradient update weights to the neuron connections corresponding to high-frequency features to make them converge faster during training; and reducing the weight of non-core features through random masking technology to reduce the interference of redundant information on decision-making.
[0079] ⑥ Gradient penalty mechanism:
[0080] When the main model incorrectly rejects the correct answer, and the meta-model confirms the error through adversarial examples, the penalty strength is calculated based on the confidence level of the main model when it incorrectly rejects the answer. The higher the penalty strength, the greater the parameter correction. During the backpropagation phase of the main model, the gradient direction is reversed on the neuron connections that caused the incorrect rejection. The maximum magnitude of a single gradient update is set to avoid damaging the model stability due to excessive penalty.
[0081] ⑦ Feature reconstruction mechanism:
[0082] When the main model incorrectly accepts an incorrect answer, and the meta-model detects such errors through adversarial examples, it performs cluster analysis on the semantic features of the incorrectly accepted answers to locate the confusion region in the main model's encoding layer. Based on the number of incorrect samples, it selects specific channels in the encoding layer for reconstruction in proportion. It only unfreezes the neurons in the selected channels for fine-tuning, while keeping other parameters frozen to ensure that the main structure of the knowledge base is not affected.
[0083] ⑧ Dynamic threshold mechanism:
[0084] The system dynamically calibrates the main model's judgment threshold based on the meta-model's adversarial verification results. Using 500 test samples as a window, the system statistically analyzes the confidence difference between the main model and the meta-model. If the difference persists, the system adjusts the main model's threshold by a fixed step size. A threshold fluctuation range is set to prevent extreme adjustments. A "delayed effect" mechanism is also introduced, requiring the threshold to take effect only when the trend is consistent across three consecutive monitoring windows.
[0085] d) Dynamic knowledge graph
[0086] ⑤ Coupling of the answer integrity graph and the dual-model collaborative architecture
[0087] In the initial answer matching stage of the main evaluation model, over 12,000 pre-defined answer nodes in the knowledge graph are used as a semantic integrity benchmark. Answers are mapped to vectors using graph embedding technology, and graph structure similarity constraints are introduced when calculating similarity with user responses. The meta-evaluation model learns the matching bias patterns of the main model through adversarial training, and, combined with the node coverage status of the integrity graph, verifies whether the main model has missed any key knowledge branches.
[0088] ⑥ Dynamic adaptation of feature weight map and four-dimensional feedback mechanism
[0089] Based on the industry feature weight map, a domain-sensitive feature enhancement mechanism is designed to synchronously associate the feature weight map: if the model's evaluation result in a specific domain deviates significantly from the map weight distribution, the gradient penalty coefficient of the domain features is automatically increased, forcing the model parameters to converge to the weight distribution labeled in the map.
[0090] ⑦ Optimization of dynamic threshold mechanism driven by time-sensitive personality spectrum
[0091] The daily / weekly / monthly frequency markers of the time-sensitive personality spectrum are linked to a dynamic threshold mechanism to design a time decay function. When the main model outputs results, the confidence threshold is dynamically adjusted based on the time-sensitive markers.
[0092] ⑧ Full-link closed loop of graph evolution mechanism and model parameters
[0093] A closed loop of evaluation, graph, and training is constructed. For example, after each evaluation cycle, based on the consistent judgment results of the principal / meta model, the knowledge graph node weights are automatically adjusted according to the rule of "update frequency = evaluation cycle × 0.3". Simultaneously, the feature reconstruction mechanism, based on the evolved graph, uses a graph neural network (GNN) to encode node weights and connections, generating synthetic data with enhanced domain features. This data is then injected into the adversarial training set to improve the model's adaptability to knowledge changes.
[0094] e) Reinforcement learning framework
[0095] ③ Deep integration of reward function and evaluation system
[0096] The reward mechanism achieves multi-level reward determination through deep integration with the adversarial verification process of the dual-model collaborative architecture. Dual verification, involving adversarial verification and dynamic threshold linkage, ensures the accuracy of reward signals and the rationality of exploration directions.
[0097] ④ Coordinated implementation of PPO algorithm and model architecture
[0098] The policy network employs a multimodal input design, concatenating the main model confidence, meta-model validation features, four-dimensional feedback metrics, and knowledge graph embedding vectors into a 512-dimensional feature set. Decisions are then made through a 3-layer MLP (512-256-128 nodes) with residual connections. A feature enhancement gating mechanism is introduced at the 256-node layer to dynamically adjust the information flow intensity based on the feature weights in the four-dimensional feedback. The exploration strategy uses a dynamic ε-greedy algorithm, with an initial 15% exploration rate automatically adjusted by statistically analyzing the effective exploration rate over a sliding window (100 steps). This architecture achieves real-time feedback between policy decisions and feature weights, ensuring a balance between exploration and utilization.
[0099] This invention significantly improves the adaptive capability of the evaluation system in scenarios of model iteration and knowledge expansion, and provides a reliable evaluation benchmark for the continuous optimization of artificial intelligence models.
[0100] The above description is merely a preferred embodiment of the present invention and is used only to illustrate the technical solution of the present invention, and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the scope of protection of the present invention.
Claims
1. A method for optimizing and dynamically calibrating an evaluation model based on meta-evaluation feedback, characterized in that, By working together with the main evaluation model and the meta-evaluation model, and combining dynamic knowledge graphs and reinforcement learning frameworks, an evaluation system with real-time feedback optimization capabilities is constructed. The main evaluation model is responsible for performing the original answer matching task. The meta-evaluation model achieves error correction and logical optimization through variance analysis, semantic alignment and adversarial verification. The two models form a mutual verification mechanism through adversarial training, and dynamic perturbations are injected using generative adversarial networks to improve the robustness of the system. Synergy between the main evaluation model and the meta-evaluation model (1) The main evaluation model performs the original answer matching task. The main evaluation model, as the basic execution layer, is responsible for directly processing input data and generating preliminary results. In the question-answering system, the main model filters the original answers through semantic vector matching or keyword retrieval. (2) The meta-evaluation model performs a secondary verification of the main model's judgment. The meta-evaluation model enhances the reliability of the main model through multi-dimensional validation, including: Error correction: Identify biases in the master model using analysis of variance and Kendall's coefficient of harmony method, and dynamically adjust the weights; Feedback optimization: Step-level feedback guides the iterative inference path of the main model; (3) Improve system robustness through adversarial training Adversarial training enhances a model's resistance to noise and anomalous inputs by simulating attack scenarios. This can be achieved through methods such as: Dynamic perturbation generation: By using generative adversarial networks (GANs) or projection adversarial training, small perturbations are injected into the input data, forcing the main model to learn more stable feature representations; Dual-model cross-validation mechanism: The main model and the meta-model act as adversaries during training; Robustness evaluation metrics: Combine statistical testing to quantify the performance degradation of the model in adversarial environments, and guide the adjustment of training strategies; It adopts a deep integration of a four-dimensional feedback mechanism and a dynamic knowledge graph; The four-dimensional feedback mechanism includes: Feature enhancement mechanisms optimize model convergence efficiency by enhancing high-frequency feature weights and suppressing interference. The gradient penalty mechanism uses a reverse correction strategy to correct erroneous decisions; Feature reconstruction mechanisms are used to specifically repair confused regions in the coding layer; The dynamic threshold mechanism achieves threshold safety calibration based on a statistical monitoring window; (1) Feature enhancement mechanism: When the main model correctly matches the answer and the meta-model confirms that its confidence level is reasonable, Toucht performs a feature enhancement process: extracting key features of the correct answer from the semantic encoding layer of the main model, comparing them with the feature library of historical correct samples, and identifying high-frequency activated feature dimensions; adding gradient update weights to the neuron connections corresponding to high-frequency features to make them converge faster during training; and reducing the weight of non-core features through random masking technology to reduce the interference of redundant information on decision-making. (2) Gradient penalty mechanism: When the main model incorrectly rejects the correct answer, and the meta-model confirms the error through adversarial examples, the penalty strength is calculated based on the confidence level of the main model when it incorrectly rejects the answer. The higher the penalty strength, the greater the parameter correction. During the backpropagation phase of the main model, the gradient direction is reversed on the neuron connections that caused the incorrect rejection. The maximum magnitude of a single gradient update is set to avoid damaging the model stability due to excessive penalty. (3) Feature reconstruction mechanism: When the main model incorrectly accepts an incorrect answer, and the meta-model detects such errors through adversarial examples, it performs cluster analysis on the semantic features of the incorrectly accepted answer to locate the confusion region in the main model's encoding layer; based on the number of incorrect samples, it selects specific channels in the encoding layer for reconstruction in proportion; it only unfreezes the neurons in the selected channels for fine-tuning, while keeping other parameters frozen to ensure that the main structure of the knowledge base is not affected; (4) Dynamic threshold mechanism: The main model's judgment threshold is dynamically calibrated based on the meta-model's adversarial verification results. N test samples are set as windows, and the confidence difference between the main model and the meta-model is statistically analyzed. If the difference persists, the main model threshold is adjusted by a fixed step size. A threshold fluctuation range is set to prevent extreme adjustments. At the same time, a delayed effect mechanism is introduced, which only takes effect when the trend is consistent for three consecutive monitoring windows. Dynamic knowledge graph (1) Coupling of answer integrity graph and dual-model collaborative architecture In the original answer matching stage of the main evaluation model, 12,000+ pre-set answer nodes in the knowledge graph are used as the semantic integrity benchmark library; the answers are mapped to vectors through graph embedding technology, and graph structure similarity constraints are introduced when calculating similarity with user answers; the meta-evaluation model learns the matching deviation pattern of the main model through adversarial training, and verifies whether the main model has missed key knowledge branches by combining the node coverage status of the integrity graph. (2) Dynamic adaptation of feature weight map and four-dimensional feedback mechanism Based on the industry feature weight map, a domain-sensitive feature enhancement mechanism is designed to synchronously associate the feature weight map: if the model's evaluation result in a specific domain deviates from the map weight distribution, the gradient penalty coefficient of the domain features is automatically increased to force the model parameters to converge to the weight distribution labeled in the map. (3) Optimization of dynamic threshold mechanism driven by time-sensitive personality spectrum The daily / weekly / monthly frequency markers of the time-sensitive personality spectrum are linked with a dynamic threshold mechanism to design a time decay function; when the main model outputs results, the confidence threshold is dynamically adjusted in conjunction with the time-sensitive markers. (4) The whole-link closed loop of the map evolution mechanism and model parameters An evaluation-graph-training closed loop is constructed. Meanwhile, the feature reconstruction mechanism is based on the evolved graph and uses a graph neural network (GNN) to encode node weights and connections, generating synthetic data with enhanced domain features. This data is then injected into the adversarial training set to improve the model's adaptability to knowledge changes.
2. The method according to claim 1, characterized in that, The dynamic knowledge graph, through the integrity constraints of 12,000+ answer nodes, industry feature weighting, and timeliness labeling, forms a closed-loop linkage with the evaluation system to generate synthetic data and enhance the knowledge evolution and adaptability.
3. The method according to claim 1, characterized in that, The reinforcement learning framework employs the PPO algorithm with multimodal input to achieve real-time feedback between the evaluation strategy and the knowledge graph.
4. The method according to claim 3, characterized in that, Reinforcement learning framework (1) Deep integration of reward function and evaluation system The reward mechanism achieves multi-level reward determination through deep integration with the adversarial verification process of the dual-model collaborative architecture; and ensures the accuracy of reward signals and the rationality of exploration directions through dual verification of adversarial verification and dynamic threshold linkage. (2) Cooperative implementation of PPO algorithm and model architecture The policy network adopts a multimodal input design, concatenating the main model confidence, meta-model validation features, four-dimensional feedback indicators, and knowledge graph embedding vectors into a 512-dimensional feature. Decisions are made through a 3-layer MLP with residual connections, i.e., 512-256-128 nodes. A feature enhancement gating mechanism is introduced at the 256-node layer to dynamically adjust the information flow intensity based on the feature weights in the four-dimensional feedback. The exploration strategy adopts a dynamic ε-greedy algorithm, with an initial 15% exploration rate that is automatically adjusted by statistically analyzing the effective exploration rate through a sliding window.