A large-scale model ensemble decision-making method based on multidimensional evidence reasoning and dynamic negotiation mechanism

By employing multidimensional evidence reasoning and dynamic negotiation mechanisms, the problems of illusion and insufficient domain adaptability in generative large models are resolved, resulting in highly credible and interpretable predictions and improving the robustness of large model ensemble decision-making.

CN123117337APending Publication Date: 2026-11-17HARBIN INST OF TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611319956.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-28
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing generative large models suffer from illusion problems in prediction, insufficient domain adaptability, low decision credibility and lack of interpretability. Traditional model ensemble methods cannot dynamically adjust dependencies, resulting in limited overall performance.

Method used

Through multi-dimensional evidence reasoning and dynamic negotiation mechanisms, multi-modal evidence extraction and probability standardization are performed to generate structured reasoning texts, assess uncertainty and sample complexity, achieve negotiated fusion and output, and dynamically adjust model weights.

Benefits of technology

It improves the credibility and interpretability of prediction results, suppresses the risk of hallucination, dynamically adapts to sample difficulty, and enhances the robustness and reliability of the system.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123117337A_ABST
    Figure CN123117337A_ABST
Patent Text Reader

Abstract

A large-scale model ensemble decision-making method based on multi-dimensional evidence reasoning and dynamic negotiation mechanism is disclosed. The method includes the following steps: Step S1, multi-modal evidence extraction and probability output; Step S2, reason text feature vector extraction and uncertainty assessment; Step S3, sample complexity assessment; Step S4, negotiated fusion and output. This invention standardizes the predicted probabilities output by each base model and generates structured prediction reasons. It uses a pre-trained language model to vectorize the reason text, quantifies model uncertainty using indicators such as entropy, and suppresses the illusionary risk of "reasons not supporting labels" through label-reason semantic consistency calibration features. Furthermore, it calculates sample complexity features based on multi-model prediction divergence to characterize the inherent classification difficulty of the samples. Finally, it inputs multi-dimensional features such as reason semantic vectors, uncertainty, consistency calibration, and sample complexity into an attention negotiation network to dynamically generate fusion weights for each model. The probability distributions of each model are then weighted and fused to output the final predicted label and comprehensive confidence score. This invention can adaptively adjust the model's discourse power on samples of varying difficulty, improving the robustness, accuracy, and interpretability of integrated decision-making, and providing a feasible intelligent negotiation and fusion solution for multi-model joint prediction.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and multimodal deep learning technology, specifically to a large-model ensemble decision-making method based on multidimensional evidence reasoning and dynamic negotiation mechanisms. It comprises four parts: first, multimodal evidence extraction and probability output; second, reasoning text feature vector extraction and uncertainty assessment; third, sample complexity assessment; and fourth, negotiated fusion and output. Background Technology

[0002] Currently, generative large models, with their deep logical deduction and knowledge transfer capabilities, are being widely applied in various predictive tasks, such as sequence labeling and multi-classification tasks. The application of large models significantly reduces the cost of manually processing unstructured data, and, thanks to their generalized knowledge learned from massive corpora, they demonstrate greater convenience and efficiency than traditional algorithm models in feature extraction and pattern recognition in complex data environments.

[0003] However, despite the excellent performance of generative large models, relying on a single model for prediction still faces many technical bottlenecks and challenges. First, large models themselves suffer from the "illusion" problem; their predicted labels and the generated rationale text are often semantically inconsistent, leading to low decision credibility. Simultaneously, general-purpose large models often exhibit cognitive biases due to differences in training data distribution when facing specific professional domains, resulting in insufficient domain adaptability. Second, traditional model ensemble methods rely solely on the numerical fusion of output probabilities, completely ignoring the logical basis behind model decisions, leading to a lack of interpretability and difficulty in identifying high-confidence errors. Furthermore, these methods typically employ static fusion weights, failing to dynamically adjust the dependencies between models based on the inherent complexity of the samples. This results in rigid strategies when facing samples of varying difficulty, limiting overall performance.

[0004] To address the above problems, this invention proposes a large-model ensemble decision-making method based on multi-dimensional evidence reasoning and a dynamic negotiation mechanism. First, this invention performs multi-modal evidence extraction and probability standardization to eliminate output scale differences between different modalities and improve the accuracy of probability estimation. Further, through reason text generation and vectorized encoding, each model is forced to output structured prediction reason text and convert it into high-dimensional feature vectors, thus making the implicit reasoning process explicit as a computable evidence representation. Subsequently, this invention achieves precise quantification and dynamic weight allocation of prediction credibility and learning difficulty through an uncertainty and sample complexity evaluation mechanism. Finally, through a negotiated fusion and output strategy, intelligent negotiation and joint decision-making among multiple models are achieved, thereby mitigating the prediction bias of a single model and improving the credibility of the prediction results.

[0005] To address the problems existing in the current technology, this invention proposes a large-model ensemble decision-making method based on multidimensional evidence reasoning and dynamic negotiation mechanisms. On the one hand, this invention automatically identifies and suppresses the illusionary risk of large models by introducing a consistency calibration mechanism between reasons and labels. On the other hand, this method combines sample complexity assessment based on prediction disagreement, enabling the system to achieve dynamic adaptive fusion. This allows for efficient consensus-building on simple samples while prioritizing models with strong logical consistency on complex samples. Furthermore, this scheme deeply integrates the model's reasoning logic to enhance the interpretability of decisions and significantly improves the robustness and reliability of the overall system through this multidimensional, adaptive negotiation mechanism.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] A large-scale model ensemble decision-making method based on multidimensional evidence reasoning and dynamic negotiation mechanism includes the following steps:

[0008] Step S1: Multimodal evidence extraction and probability output

[0009] The aim of this step is to unify the output format of multiple models and to express the implicit derivation process explicitly in text, providing a standardized input basis for subsequent evidence fusion. To this end, this invention provides a multimodal evidence extraction method to extract multimodal data, including numerical and textual data, from the prediction processes of different models.

[0010] (1) Extracting standardized prediction probabilities: This invention converts the original prediction results of each model into a one-hot coding mode, calculates the corresponding objective probability value for each label, and normalizes the probability vector to ensure that the standardized probability distribution can be compared and fused in a unified probability space.

[0011] (2) Generating Structured Prediction Reasons: This invention obtains the corresponding structured reason text for each model in the process of outputting prediction results by constructing a prompt word framework for extracting reason text. Among them, the reason text needs to follow the preset format constraints and include key information such as evidence points, reasoning logic and uncertainty explanation, so as to facilitate subsequent vectorization encoding and evidence analysis processing.

[0012] Step S2: Extraction of feature vectors from the reasoning text and assessment of uncertainty

[0013] This step aims to convert the structured reasoning texts generated by each modality into quantifiable feature representations and assess the degree of uncertainty in the model's prediction process, providing a reliable quantitative basis for subsequent evidence fusion and decision-making. To this end, this invention provides a deep learning-based method for text feature extraction and uncertainty measurement to fully extract semantic information and confidence signals from the reasoning texts.

[0014] (1) Vectorization of reason text features: This invention uses a pre-trained BERT model to encode the structured reason text generated by each model, and extracts the hidden state corresponding to the [CLS] label of the last layer of BERT as the overall feature vector representation of the reason text, ensuring that each reason text can be converted into a dense vector of uniform dimension.

[0015] (2) Uncertainty Tendency Assessment: This invention assesses the uncertainty of each model's predictions through methods such as entropy calculation of probability distribution, marginal difference analysis, and self-consistent sampling verification. Among them, entropy calculation mainly measures the dispersion of the predicted probability distribution, marginal difference analysis mainly assesses the gap between the highest probability and the second highest probability, and self-consistent sampling verifies the stability of the prediction results through multiple generations. These indicators are combined to form an uncertainty tendency score for the generation rationale of each model.

[0016] (3) Consistency calibration feature calculation: This invention maps the predicted label and the generated reason to the semantic space of the same pre-trained language model, and calculates the cosine similarity of their semantic vectors as a consistency calibration feature. High similarity indicates that the reason fully supports the prediction and is logically consistent; low similarity reveals semantic conflict between the reason and the label, and there is a risk of illusion. This feature will be used to dynamically adjust the model weights in the subsequent fusion stage, thereby reducing decision-making errors caused by logical inconsistency.

[0017] Step S3: Sample Complexity Assessment

[0018] This step identifies the difficulty characteristics of the sample by evaluating its complexity attributes. The core purpose of this step is to examine the differences in performance of different models on samples of varying complexity, and to use the sample complexity evaluation results as an important basis for guiding the weight allocation of each modality model.

[0019] This invention first aggregates the predicted probability distributions of all base models for the current sample, calculating an average probability distribution that represents the "group consensus." Then, it quantifies the divergence by measuring the difference between each model's predicted distribution and this average consensus distribution. The average divergence of all models is calculated as the final sample complexity feature: a higher value indicates greater divergence between models, making the sample more ambiguous and difficult to classify; a lower value indicates clearer sample features and lower classification difficulty.

[0020] Step S4: Negotiated Integration and Output

[0021] This step aims to integrate the multimodal evidence information obtained in the preceding steps, and to achieve negotiation and joint decision-making among the models through deep learning algorithms, ultimately generating stable and reliable integrated prediction results. This invention constructs a negotiation-based fusion architecture based on a combination of attention mechanisms and pooling strategies, achieving adaptive weight allocation and intelligent fusion of multimodal evidence.

[0022] (1) Construction of multi-dimensional feature fusion vector: For each base model, this invention concatenates the rationale text vector, uncertainty score, and consistency calibration features of the model, and additionally concatenates the global sample complexity features, so that the subsequent fusion module can simultaneously perceive the model quality and sample difficulty. If the scale difference between the scalar features and the high-dimensional text vector is large, a lightweight mapping layer is first used to compress the scalar to a suitable dimension before concatenation, so as to ensure that the feature fusion is stable and effective.

[0023] (2) Adaptive negotiation weight calculation: The present invention uses an attention mechanism to input the comprehensive feature vector of each model into the attention network, outputs the trust score of the model, and obtains its fusion weight in the final decision through normalization.

[0024] (3) Weighted decision fusion and result output: After obtaining the negotiated weights of each base model, this invention performs weighted summation of the category probability distributions output by each model to obtain the final probability distribution after global fusion, and selects the category with the highest probability as the final predicted label. At the same time, the corresponding comprehensive confidence score is output, which has stronger robustness and interpretability compared with single model or simple average fusion.

[0025] Compared with existing methods, the present invention has the following advantages:

[0026] (1) Suppressing the risk of hallucination: This invention introduces a consistency calibration feature that can automatically reduce the model weights of those "reasons that do not support the label".

[0027] (2) Strong dynamic adaptability: This invention can quickly reach a consensus on simple samples through sample complexity evaluation, and focus on more logically consistent models on complex samples.

[0028] (3) Balance between interpretability and robustness: The fusion process of this invention not only relies on probability values, but also deeply understands the reasoning logic of the model, realizing the leap from "numerical weighting" to "logical negotiation". Attached Figure Description

[0029] Figure 1 This is a flowchart of a large-scale integrated decision-making method based on multidimensional evidence reasoning and dynamic negotiation mechanisms. Detailed Implementation

[0030] The technical solution of the present invention will be further described below with reference to the accompanying drawings, but it is not limited thereto. Any modifications or equivalent substitutions to the technical solution of the present invention that do not depart from the spirit and scope of the technical solution of the present invention should be covered within the protection scope of the present invention.

[0031] This invention provides a large model ensemble method based on multimodal evidence, such as... Figure 1 As shown, the method includes the following steps.

[0032] Step S1: Multimodal evidence extraction and probability output

[0033] (1) Extracting standardized prediction probabilities

[0034] First, assume there are N large models performing classification on C categories, where the original output of the i-th model is... Furthermore, each model outputs the predicted probability for each category as shown in Equation (1).

[0035]

[0036] in, Let represent the predicted probability of the i-th model for category c.

[0037] Then, the predicted probability of each model under each category is normalized to ensure that the prediction results of different models can be compared under the same probability space, as shown in Equation (2).

[0038]

[0039] The probability distribution after normalization must satisfy equation (3).

[0040]

[0041] (2) Generate structured reasons

[0042] After obtaining standardized prediction probabilities, this invention further requires each model to output an explicit reasoning process. Since the output of generative large models usually has the characteristics of free text and is difficult to directly perform numerical calculations, this step constructs a standardized prompt template framework so that the model outputs prediction reasons according to a preset structure, as shown in Table 1.

[0043] Table 1. Structured Reason Generation Prompt Templates

[0044] Module Name Description of prompt words Example of specific instructions Character settings Assign the role of a model expert and activate domain knowledge. You are an experienced data annotation expert and logic analyst. Task Instructions Define the input content and prediction target Please read the following text to be analyzed, determine its category, and analyze the key evidence leading to this conclusion. Input slot Dynamically fill in the samples to be predicted [Text to be analyzed]: {input_text} Format constraints Force output to JSON or a specific delimiter format to facilitate subsequent cleaning. Please strictly adhere to the following JSON format when outputting results, and do not include any additional explanations: json { "predicted_label": "Predicted category name", "rationale": A concise and clear reasoning statement, including key entities or logical chains, with a length of no more than 50 words. }

[0045] After the model generates a response, this invention uses an automated script to clean and parse the data. First, the system attempts to parse the model's original output text as a standard JSON object. If the model strictly adheres to the aforementioned format constraints, the text corresponding to the `rationale` field is directly extracted as the structured reason. Second, this invention uses regular expressions for error-tolerant extraction. By configuring the matching mode, it ignores newline characters and spaces, accurately locating and extracting the reason text within double quotes. Finally, the extracted text is normalized, removing newline characters, escape characters, and meaningless stop words. If the extracted content is empty or parsing fails completely, the system automatically fills in a default marker (such as [UNK]) to ensure consistency in the input dimensions of subsequent feature vectorization steps.

[0046] Step S2: Extraction of feature vectors from the reasoning text and assessment of uncertainty

[0047] (1) Reason text feature vectorization

[0048] This step aims to transform the unstructured reason text extracted in step S1 into a high-dimensional semantic vector that can be computed by a computer. First, a pre-trained language model is loaded as a feature extractor. The cleaned reason text is input into the model's tokenizer, and a start token [CLS] and an end token [SEP] are added, and the tokens are truncated or padded to a fixed length. Subsequently, the processed token sequence is input into the Transformer encoding layer of the pre-trained model. In this invention, the hidden state vector corresponding to the [CLS] token in the output of the last layer of the model is selected as the global semantic representation of the reason text, and the specific calculation process is shown in Equation (4).

[0049] in, The dimension d is usually 768 dimensions.

[0050] (2) Uncertainty Tendency Assessment

[0051] This step quantifies the confidence level of each base model in predicting the current sample, using this as a reference for subsequent fusion weights. It utilizes the probability distribution normalized by Softmax in step S1. The information entropy of this distribution is calculated as an uncertainty index. The higher the information entropy value, the flatter the probability distribution, and the more hesitant the model is in classifying the sample; conversely, the lower the entropy value, the more pronounced the model's bias. The formula for calculating the uncertainty score H(P) is shown in equation (5).

[0052]

[0053] Where C represents the total number of categories in the classification task. This represents the probability value that the model predicts a sample belongs to class c. To facilitate subsequent neural network processing, this invention further standardizes the calculated entropy value, mapping it to the [0,1] interval, thereby forming a scalar feature that matches the text vector. .

[0054] (3) Calculation of consistency calibration characteristics

[0055] To quantify whether there is a semantic and logical conflict between the text reasoning generated by the detection model and the predicted label, i.e., the phenomenon of "inconsistency between words and deeds," and to identify potential risks of large model illusion, this invention first inputs the category label name predicted by the current model into the same pre-trained language model as in step S2 (1), extracts the hidden state at the [CLS] position, and obtains the label semantic vector. This ensures that the label and the reason are compared within the same semantic space. Secondly, the cosine similarity function is used to calculate the feature vector of the reason. With label semantic vector The degree of matching between them. This calculation process measures the support of the reason text for the predicted label in the semantic direction, and the calculation formula is shown in Equation (6).

[0056]

[0057] Where · represents the vector dot product, This represents the magnitude of the vector. Finally, the calculated similarity score is... As a consistency calibration feature. If A value close to 1 indicates that the rationale behind the model's generation strongly supports its predictions and demonstrates strong logical consistency; if A low or even negative value indicates that although the model provides a classification label, the reason for its generation is irrelevant or contradictory, and the weight of the sample should be reduced in subsequent fusion.

[0058] Step S3: Sample Complexity Assessment

[0059] This step aims to quantify the inherent difficulty of semantic understanding and classification of the current input sample. Generally, if multiple base models show high consistency in their predictions for the same sample, it indicates that the sample has obvious features and low complexity; conversely, if there are significant discrepancies between the models, it means that the sample has high ambiguity or complexity. First, the system aggregates the probability distributions predicted by K base models for the same input sample. Each of them All vectors are Softmax normalized. To obtain the "consensus" distribution of the model population, the arithmetic mean of the prediction probabilities of all models is calculated. The calculation formula is shown in (7).

[0060] (7)

[0061] Secondly, the Jensen-Shannon divergence is used to measure the probability distribution of each pedestal model. Compared with the average distribution The degree of difference between them. Sample complexity features. The calculation is obtained by taking the average value of the KL divergence of each model, as shown in equation (8).

[0062] (8)

[0063] in, This represents the KL divergence, where C is the total number of categories. This represents the probability that the k-th model predicts a sample belonging to the c-th class. This represents the probability of an evenly distributed distribution in class c. Finally, the calculated scalar... This value serves as a feature of sample complexity. The larger the value, the more significant the divergence between the base models, and the higher the difficulty of sample classification.

[0064] Step S4: Negotiated Integration and Output

[0065] (1) Construction of multidimensional feature fusion vector

[0066] This step aims to construct a comprehensive feature representation for each base model, including its decision logic and confidence state, for evaluation by the meta-learner. For the k-th base model, the reason text feature vector extracted in step S2 is used. Uncertainty score and consistency calibration features The data is then stitched together. Simultaneously, to enable the fusion network to perceive the current sample difficulty, the global sample complexity feature calculated in step S3 is used. This is then concatenated into the feature vector of each model. The resulting composite feature vector... The formal expression is shown in equation (9).

[0067] (9)

[0068] in, This represents a vector concatenation operation. If the scalar features and the high-dimensional text vectors differ significantly in dimensionality, the scalars can be mapped to a low-dimensional space using a fully connected layer before concatenation.

[0069] (2) Adaptive negotiation weight calculation

[0070] This step utilizes an attention mechanism to simulate the "negotiation" process among multiple models, dynamically calculating their influence in the final decision based on their overall performance. The constructed feature vectors... The input is fed into an attention network composed of neural networks to calculate the nonnormalized trust score for each pedestal model. The network learns to identify which models provide logically consistent and highly confident evidence. Then, the scores are normalized using the Softmax function to obtain the final fusion weights for each pedestal model k. The specific formulas are shown in (10) and (11).

[0071]

[0072]

[0073] Where W and b are the learnable weight matrix and bias term of the attention network, respectively, and K is the total number of base models.

[0074] (1) Weighted decision fusion and result output

[0075] This step generates the final classification result based on the weights determined through negotiation. The weight coefficients obtained in step (2) are used... The original probability distribution output for all base models Perform a weighted summation to generate the globally optimal probability distribution. Finally, select The category with the highest probability value is used as the final predicted label. The entire fusion decision-making process can be represented by equations (12) and (13).

[0076]

[0077]

[0078] The system ultimately outputs predicted labels. And the corresponding comprehensive probability confidence level. Since this result is derived based on reason semantics and multi-dimensional quantitative indicators, it has higher robustness and interpretability than a single model.

[0079] It should be noted that the effective implementation of this invention relies on several preconditions and engineering constraints. First, the task requires pre-defining and solidifying the label system, forming a clear set of categories and their standard names / codes, and ensuring that the output probability distribution and label predictions of each base model are strictly aligned to the same label space. Second, to improve the stability of consistency checks, it is recommended to configure standardized label text for each label to avoid semantic similarity deviations caused by different models using similar but inconsistent label expressions. Third, this solution requires each model to output structured reasoning text according to a preset prompt, which must be parsable, and fault tolerance and default placeholders should be set for parsing failures to ensure consistent vectorized input dimensions. Finally, attention fusion networks typically need to be trained or calibrated on labeled data to learn the mapping relationship between "reasoning consistency / uncertainty / sample complexity" and the final weight allocation; simultaneously, the output probabilities of each base model should be uniformly calibrated to avoid fusion bias caused by differences in model probability scales.

Claims

1. A large-scale model ensemble decision-making method based on multidimensional evidence reasoning and dynamic negotiation mechanism, characterized in that, Includes the following steps: Step S1, Multimodal Evidence Extraction and Probability Output: Obtain the sample to be predicted, call K base models to reason about the sample to be predicted, obtain the category prediction probability distribution of each base model, and obtain the reason text corresponding to the prediction results of each base model. Step S2, Extraction of Reason Text Feature Vectors and Uncertainty Assessment: The reason text is vectorized to obtain the reason text feature vectors of each base model; the uncertainty features of each base model are calculated based on the category prediction probability distribution. And calculate the consistency calibration features between the predicted labels of each pedestal model and the reason text; Step S3, Sample Complexity Evaluation: Based on the differences between the category prediction probability distributions of the K base models, calculate the sample complexity characteristics of the sample to be predicted; Step S4, Negotiated Fusion and Output: For each base model, construct a comprehensive feature representation that includes the reason text feature vector, uncertainty feature, consistency calibration feature, and sample complexity feature. Input the comprehensive feature representation into the negotiated fusion model to obtain the fusion weight of each base model. Then, perform weighted fusion on the category prediction probability distribution of each base model according to the fusion weight, and output the final predicted label and / or the final probability distribution.

2. The method according to claim 1, characterized in that, In step S1, the category prediction probability distribution output by each base model is unified. The unification process includes any one or more of probability normalization, probability calibration, or scale alignment, so that the outputs of different base models are in a comparable probability space.

3. The method according to claim 1, characterized in that, The reason text mentioned in step S1 is constrained to a structured output by the prompt word template. The structured output is any one or more of JSON, key-value pairs, or delimiter format, and includes at least a prediction label field and a reason field.

4. The method according to claim 1, characterized in that, In step S1, when the reason text is missing or parsing fails, the reason text is completed using a default tag or default text to ensure the consistency of subsequent vectorized input.

5. The method according to claim 1, characterized in that, The reason text feature vector mentioned in step S2 is obtained by encoding the reason text through a pre-trained representation model. The pre-trained representation model includes any one or more of a bidirectional encoder model, a sentence vector model, or a generative language model.

6. The method according to claim 1, characterized in that, The uncertainty features mentioned in step S2 include any one or more of the following: information entropy based on the probability distribution of category prediction, the difference between the maximum probability and the second-highest probability, the marginal difference index, and the consistency index based on repeated generation or repeated sampling.

7. The method according to claim 1, characterized in that, The consistency calibration feature mentioned in step S2 is obtained by calculating the similarity after mapping the predicted label and the reason text to the same semantic space. The similarity includes any one or more of cosine similarity, dot product similarity, distance metric similarity, or learnable matching function.

8. The method according to claim 1, characterized in that, The sample complexity feature mentioned in step S3 is obtained by characterizing the prediction divergence between the base models. The prediction divergence is based on any one or more of Jensen–Shannon divergence, Kullback–Leibler divergence, total variation distance, Wasserstein distance, or divergence degree based on ordination difference.

9. The method according to claim 1, characterized in that, The comprehensive features mentioned in step S4 are represented as vector concatenation, or are obtained by mapping scalar features into vectors through a mapping network and then fusing them with the reason text feature vectors; wherein the scalar features include at least two of the following: uncertainty features, consistency calibration features, and sample complexity features.

10. The method according to claim 1, characterized in that, The negotiation fusion model mentioned in step S4 includes any one or more of the attention mechanism model, gating network model, hybrid expert model or meta-learning model, and the fusion weights are normalized to make the sum of the fusion weights of each base model equal to 1.

11. The method according to claim 1, characterized in that, The weighted fusion described in step S4 involves summing the category prediction probability distributions of each base model according to the fusion weights, and using the category with the highest probability in the fused probability distribution as the final predicted label.

12. The method according to claim 1, characterized in that, In step S4, a comprehensive confidence score is further output. The comprehensive confidence score is determined by the peak value of the fused probability distribution, the fused distribution entropy, the fused weight concentration, or a combination thereof.

13. A large-scale model integrated decision-making device, characterized in that, include: The multimodal evidence extraction and probability output module is used to perform step S1 as described in claim 1; the feature vector extraction and evaluation module is used to perform step S2 as described in claim 1; the sample complexity evaluation module is used to perform step S3 as described in claim 1; and the negotiated fusion and output module is used to perform step S4 as described in claim 1.

14. A computer device, characterized in that, The method includes a processor and a memory, wherein the memory stores a computer program that, when executed by the processor, implements the method according to any one of claims 1 to 12.

15. A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method of any one of claims 1 to 12.