A bias detection method and system for mental illness large model evaluation

Abstract

The application discloses a mental illness large model evaluation bias detection method and system, and relates to the technical field of mental health intelligent evaluation. The method comprises the following steps: using a self-preference bias coefficient to quantize the bias of review results based on homologous model output and heterologous model output according to a fine-grained evaluation criterion; structurally analyzing and risk-identifying bias-induced conditions according to fine-grained evaluation criterion attributes; identifying, isolating and correcting the bias of review results based on a multi-model heterogeneous consensus detection mechanism; and bias-correcting the review results according to the bias quantization results, the criterion attribute risk identification results and the consistency score to obtain the final bias detection results. The application can realize fine-grained description, cross-criterion tracking, quantitative analysis and structural attribution of the bias in the mental illness large model criterion evaluation.

Description

Technical Field

[0001] This invention relates to the field of intelligent assessment technology for mental health, and in particular to a method and system for detecting biases in the assessment of a large model of mental illness. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] To overcome the shortcomings of traditional large-scale mental illness model assessments in terms of interpretability and stability, large-scale mental illness model assessments are gradually shifting towards a criterion-based assessment system, namely criterion-based evaluation (RB). However, existing research still lacks a systematic understanding of the biases inherent in large-scale mental illness models under the criterion-based assessment paradigm. While criterion constraints may formally regulate assessment behavior, they cannot inherently guarantee that the assessment model is completely free from potential biases. In practical applications, large-scale mental illness models may still, through more subtle and decentralized methods, give more lenient and favorable judgments to the outputs of similar models on various criterions, or accumulate an overall bias advantage across multiple local scores. Especially in mental illness scenarios, models may show higher identification with risk warning patterns, empathetic wording, conservative advice templates, or symptom reasoning paths similar to themselves, thus causing automated assessment conclusions to deviate from true clinical consensus and expert standards. Summary of the Invention

[0004] To address the shortcomings of existing technologies, the purpose of this invention is to provide a method and system for bias detection in the assessment of large-scale models of mental illness, which can achieve fine-grained characterization of biases in the assessment of large-scale models of mental illness, cross-criterion tracking, quantitative analysis, and structured attribution.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solution: The first aspect of this invention provides a method for bias detection in the assessment of a large model of mental illness, comprising the following steps: The fine-grained evaluation criteria and review results of the large model of the mental illness to be detected are obtained. The self-preference bias coefficient is used to quantify the bias of the review results based on the output of the same model and the output of the different models according to the fine-grained evaluation criteria, so as to obtain the bias quantification result. Based on the fine-grained evaluation criteria attributes, a structured analysis and risk identification of deviation inducing conditions are performed to obtain the risk identification results of the criteria attributes; Based on a multi-model heterogeneous consensus detection mechanism, biases in the review results are identified, isolated, and corrected to obtain a consistency score. Based on the deviation quantification results, the criterion attribute risk identification results, and the consistency score, the review results are corrected for deviations, and the final deviation detection results are obtained.

[0006] Furthermore, at least one review model is invoked to review multiple candidate outputs corresponding to the input sample under each fine-grained criterion, resulting in a structured review result.

[0007] Furthermore, using the self-preference bias coefficient and fine-grained evaluation criteria, the bias of the review results based on the output of the same-source model and the output of the different-source model is quantified. The specific steps to obtain the bias quantification results are as follows: Construct a self-preference bias coefficient based on the relative evaluation bias coefficient index; Using the self-preference bias coefficient based on fine-grained mental illness assessment criteria, we conducted statistical analysis on the judgment results of the review model on candidate outputs, and calculated the false pass rate, criterion tolerance, and local bias intensity on failed samples.

[0008] The order pair consistency between the output of the review model and the benchmark truth is compared to form a deviation measure index that reflects the global preference premium.

[0009] Furthermore, based on the fine-grained evaluation criterion attributes, the specific steps for performing structured analysis and risk identification of deviation-inducing conditions to obtain the criterion attribute risk identification results are as follows: Extract and record various feature information of each fine-grained evaluation criterion. Specifically, perform attribute parsing on the fine-grained evaluation criterion, set structured analysis intervals for criterion length, automatically identify the semantic polarity of the criterion, and classify the criterion topics by domain. Based on the feature information, a correlation matrix between the criterion attributes and the deviation intensity between the deviation factor and the deviation response is constructed to obtain the deviation risk level, interpretation space level and recommended handling strategy corresponding to the fine-grained evaluation criteria.

[0010] Furthermore, based on the multi-model heterogeneous consensus detection mechanism, the specific steps for identifying, isolating, and correcting biases in the review results to obtain the consistency score are as follows: Multiple review models from different technical approaches, training corpus sources, alignment methods, and parameter scales are introduced to form a review committee. Through parallel judgment and consistency analysis, the bias risks in a single model are identified. Specifically, based on the homology between the review model and the candidate output generation model, model family relationship, historical bias records, or current abnormal bias in the review, the corresponding review model is subjected to weight reduction or voting elimination.

[0011] A second aspect of the present invention provides a bias detection system for large-scale assessment of mental illness models, comprising: The bias quantification modeling module is used to obtain the fine-grained evaluation criteria and review results of the large model of the mental illness to be detected. It uses the self-preference bias coefficient to quantify the bias of the review results based on the output of the same source model and the output of the different source models according to the fine-grained evaluation criteria, and obtains the bias quantification result. The criterion attribute correlation analysis module is used to perform structured analysis and risk identification of deviation inducing conditions based on fine-grained evaluation criterion attributes, and obtain the criterion attribute risk identification results. The multi-model consensus detection module is used to identify, isolate, and correct deviations in the review results based on the multi-model heterogeneous consensus detection mechanism, and obtain a consistency score. The evaluation result correction output module is used to correct the deviation of the review results based on the deviation quantification results, the criterion attribute risk identification results, and the consistency score, and obtain the final deviation detection results.

[0012] Furthermore, it also includes a fine-grained criterion review module, which is used to call at least one review model to review multiple candidate outputs corresponding to the input sample under each fine-grained criterion item, and obtain structured review results.

[0013] Furthermore, the deviation quantification modeling module also includes: Construct a self-preference bias coefficient based on the relative evaluation bias coefficient index; Using the self-preference bias coefficient based on fine-grained mental illness assessment criteria, we conducted statistical analysis on the judgment results of the review model on candidate outputs, and calculated the false pass rate, criterion tolerance, and local bias intensity on failed samples.

[0014] The order pair consistency between the output of the review model and the benchmark truth is compared to form a deviation measure index that reflects the global preference premium.

[0015] Furthermore, the criterion attribute correlation analysis module also includes: Extract and record various feature information of each fine-grained evaluation criterion. Specifically, perform attribute parsing on the fine-grained evaluation criterion, set structured analysis intervals for criterion length, automatically identify the semantic polarity of the criterion, and classify the criterion topics by domain. Based on the feature information, a correlation matrix between the criterion attributes and the deviation intensity between the deviation factor and the deviation response is constructed to obtain the deviation risk level, interpretation space level and recommended handling strategy corresponding to the fine-grained evaluation criteria.

[0016] Furthermore, the multi-model consensus detection module also includes: Multiple review models from different technical approaches, training corpus sources, alignment methods, and parameter scales are introduced to form a review committee. Through parallel judgment and consistency analysis, the bias risks in a single model are identified. Specifically, based on the homology between the review model and the candidate output generation model, model family relationship, historical bias records, or current abnormal bias in the review, the corresponding review model is subjected to weight reduction or voting elimination.

[0017] The above one or more technical solutions have the following beneficial effects: This invention discloses a method and system for bias detection in the assessment of a large-scale mental illness model. It aims to address key shortcomings in existing automated assessment technologies for large-scale mental illness models regarding bias identification and control. Specifically, it addresses issues such as the relatively limited means of bias quantification in current assessment systems, the difficulty in effectively capturing implicit self-bias in scenarios with fine-grained objective criteria, the lack of systematic analysis of bias causes, and the absence of implementable correction mechanisms for review results. This invention proposes a method and system for detecting self-bias in a large-scale mental illness model based on fine-grained criteria. Based on the application scenario of a large-scale mental illness model as the assessor, this invention constructs an improved Relative Assessment Bias Coefficient (RABC) to quantitatively characterize and mathematically model the systematic bias exhibited by the review model when facing outputs from both homologous and heterologous models—a bias that is "lenient towards itself and strict towards others." Furthermore, it combines factors such as criterion polarity, criterion length, semantic theme, risk level, task attributes, and scoring patterns to analyze the triggering mechanism, propagation path, and strength distribution of bias under different assessment conditions, thereby forming a multi-dimensional identification and interpretable representation of self-bias in the large-scale mental illness model. Meanwhile, this invention not only focuses on the detection of biases, but also provides a full-chain technical solution for the pre-, mid-, and post-evaluation processes of mental illness model assessments. This solution covers the entire process, from criterion design optimization, review prompts and constraints, multi-model cross-validation to post-processing correction of results. This aims to actively suppress and automatically weaken biases, ultimately improving the objectivity, consistency, stability, and credibility of large-scale mental illness model assessment results. It is particularly suitable for sensitive application scenarios with high requirements for accuracy, impartiality, and safety, such as depression risk assessment, suicidal ideation identification, mental symptom counseling, crisis intervention recommendations, medication-related prompts, ethical compliance review, and patient communication support. This allows the review conclusions given by the model to more effectively approximate real human expert consensus, clinical guideline requirements, and mental health industry standards.

[0018] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart of the bias detection method for the large-scale mental illness model assessment in Embodiment 1 of the present invention; Figure 2 This is a schematic diagram of the review process for different paradigms of the large model of mental illness in Embodiment 1 of the present invention. Detailed Implementation

[0021] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0022] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof. The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0023] As Large Language Models (LLMs) demonstrate rapidly improving capabilities in areas such as medical question answering, clinical assistance, medical text generation, disease summarization, diagnostic reasoning, risk warning, and multi-task instruction following, their application scope in mental health and mental illness auxiliary assessment is constantly expanding. The model iteration speed is continuously accelerating, and the resulting demand for model performance evaluation is showing a trend towards specialization, high frequency, and complexity. Especially in scenarios involving depressive disorders, anxiety disorders, bipolar disorder, schizophrenia, post-traumatic stress disorder, personality disorders, sleep-related mental health problems, and self-harm / suicide risk identification, large-scale mental illness models are increasingly being applied to tasks such as clinical auxiliary consultation, psychological intervention support, patient education, risk screening, and health management. Therefore, how to accurately, continuously, and cost-effectively evaluate the output quality of such models has become a key fundamental issue in the deployment of intelligent mental health systems.

[0024] Traditional assessment methods relying on subjective evaluations by psychiatrists, psychologists, or crowdsourced annotators are increasingly failing to meet the demands of rapid development, frequent updates, and continuous online validation of large-scale mental illness models due to issues such as high professional labor costs, long cycles, insufficient consistency, limited scalability, and difficulty in covering massive samples. Therefore, utilizing large language models as automated evaluators—the so-called "model-as-a-judge" (LLM-as-a-Judge)—is gradually becoming an important paradigm for evaluating large-scale mental illness models and is considered a key technology supporting the training iteration, performance comparison, reward modeling, alignment optimization, and pre-deployment quality control processes of large-scale medical models.

[0025] In existing automated assessment systems, the most common assessment methods include pairwise comparison (PwC) and direct assessment (DA). PwC typically requires the assessment model to evaluate two candidate outputs based on factors such as medical accuracy, psychiatric risk identification ability, adequacy of empathic support, safety of recommendations, ethical compliance, task completion, and linguistic appropriateness. Direct assessment, on the other hand, requires the assessment model to individually score or rank candidate outputs according to given dimensions, resulting in a structured quality evaluation. These automated assessment methods offer significant advantages over manual assessment in terms of efficiency, cost, and coverage, and are therefore widely used in scenarios such as benchmarking, version comparison, training feedback, preference data construction, and pre-deployment risk control for large-scale mental illness models.

[0026] However, as related research and applications deepen, the industry has gradually discovered that large-scale models of mental illness, as reviewers, do not inherently possess absolute objectivity. Because different large-scale models of mental illness may share similar sources of medical pre-training corpora, mental health Q&A data, clinical guideline knowledge bases, fine-tuning distributions, alignment goals, parameter structures, and even generation strategies, the evaluation models, when judging candidate answers, are often potentially influenced by their own representational space, professional expression style, psychiatric terminology preferences, risk discourse frameworks, and clinical reasoning patterns, thus exhibiting a systematic bias phenomenon. Specifically, when candidate texts are consistent with the evaluation model's own familiar or closer expressions, symptom organization structures, risk warning styles, disease course reasoning habits, or psychiatric suggestion templates, the evaluation model is more likely to give them higher ratings; this tendency may be even more pronounced when candidate answers come from the same technology family, similar training system, similar medical fine-tuning framework, or similar mental health knowledge alignment strategies as the evaluation model. This phenomenon is defined as "ego bias" in this invention. The existence of this type of self-bias directly weakens the fairness and reliability of the automatic assessment results, causing systematic distortion in the performance comparison between different large models of mental illness, which in turn affects subsequent model selection, training feedback, clinical use interpretation, and safety deployment decisions.

[0027] Rubric-based evaluation (RB) methods typically break down complex comprehensive evaluation tasks into several relatively clear, operational, and interpretable evaluation dimensions or judgment criteria. These criteria include whether symptom identification is complete, risk warnings are sufficient, diagnostic statements are prudent, whether direct substitution of doctor's diagnosis is avoided, whether crisis intervention suggestions are included, whether stigmatizing language is used, whether emotional support is provided, whether ethical boundaries for mental health are followed, whether professional assistance is suggested, and whether inappropriate medication recommendations are avoided. The evaluation model is required to make binary judgments, multi-level judgments, or sub-item scores for each criterion.

[0028] Compared to a single overall preference or coarse-grained score, criterion-based assessment methods can theoretically reduce vague subjective judgments through clearer decision-making criteria, and improve the transparency, traceability, and analyzability of the mental illness model assessment process. Therefore, it is considered an important direction for building a highly reliable intelligent assessment system for mental health.

[0029] In summary, for the scenario of fine-grained criterion-based assessment of large-scale mental illness models, there is an urgent need for a technical solution capable of accurately quantifying and detecting self-bias, explicitly modeling features, performing systematic attribution analysis, and providing stable intervention. This solution should not only cover different assessment paradigms such as pairwise comparisons, direct scoring, and criterion-based evaluations, but also provide a unified analysis and processing of biases that is interpretable, transferable, and reusable, even under conditions involving multiple models, cross-source candidate answer inputs, and complex dimensions of mental health assessment. Based on this, this invention discloses a bias detection method and system for assessing large-scale mental illness models. This method effectively identifies and eliminates assessment biases, improves the objectivity of judgments, and provides more robust, fair, and reliable technical support for performance evaluation, model alignment, training feedback, reward modeling, and the construction of trustworthy medical artificial intelligence infrastructure for large-scale mental illness models. Specific details are provided in the following embodiments.

[0030] Example 1: Embodiment 1 of the present invention provides a method for bias detection in the assessment of a large model of mental illness, such as... Figure 1 As shown, it includes the following steps: S1: Call at least one review model to review multiple candidate outputs corresponding to the input sample under each fine-grained criterion to obtain structured review results.

[0031] In this invention, the review model refers to a large-scale mental illness model used for automated judgment, scoring, ranking, or reasoning generation of candidate outputs; it can be the model to be tested itself, a model with a homology relationship to the model to be tested, or a heterogeneous review unit used for cross-validation. Candidate outputs refer to the text results to be reviewed, generated by one or more models to be evaluated and submitted to the review model for judgment, for the same input sample. These include responses to mental illness questions and answers, risk warnings, consultation suggestions, explanations of diagnostic caution, crisis intervention suggestions, or ethical compliance statements, etc.

[0032] In one specific implementation, the primary step of this embodiment is to construct a unified benchmark coordinate system for the review paradigm of mental illness models, so as to conduct comparable, transferable, and quantifiable analysis of the biased behaviors generated when the large mental illness model acts as a reviewer under different automated assessment modes. For example... Figure 2As shown, the evaluation paradigms supported in this embodiment include Pairwise Comparison (PWC), Direct Assessment (DA), and Rubric-based Evaluation (RB). The PWC model is primarily used to compare the relative merits of two candidate answers in a mental illness scenario. The DA model is mainly used to assess the overall quality of a single answer. The RB model, by breaking down the overall evaluation task into multiple specific, verifiable, and executable fine-grained criteria, allows the review model to make a binary decision of pass / fail for each criterion, thereby significantly reducing the uncertainty caused by excessive subjective interpretation in traditional overall scoring methods.

[0033] In this embodiment, the RB model is used as a key application in highly sensitive scenarios such as mental illness Q&A, psychological crisis response, depression and anxiety counseling, psychiatric screening recommendations, clinical communication support, self-harm and suicide risk warnings, and medical ethics review. For example, in mental illness counseling tasks, the review model does not need to give a vague evaluation of "overall response quality of 8 points," but rather makes binary decisions on each criterion, such as "whether it clearly indicates that it cannot replace a professional psychiatric diagnosis," "whether it identifies risk signals of persistent insomnia or despair," "whether it recommends immediately contacting family members, the emergency room, or a crisis intervention hotline when self-harming thoughts occur," "whether it avoids using stigmatizing language," "whether it does not directly make a definitive diagnosis of mental illness," "whether it provides basic supportive advice rather than simple comfort," and "whether it does not give unauthorized medication adjustment suggestions." In this way, the abstract mental illness model assessment task can be mapped to a structured set of criteria, providing a stable and interpretable data foundation for subsequent bias analysis.

[0034] S2: Obtain the fine-grained evaluation criteria and review results of the large model of the mental illness to be tested. Utilize the self-preference bias coefficient to quantify the bias of the review results based on the output of the same-source model and the output of the different-source model according to the fine-grained evaluation criteria, and obtain the bias quantification result.

[0035] S2.1: Construct a self-preference bias coefficient based on the relative evaluation bias coefficient index.

[0036] In one specific implementation, this embodiment constructs a hierarchical bias calculation framework to quantitatively characterize the self-bias generated when the large model of mental illness acts as a reviewer from different analytical granularities.

[0037] To quantitatively describe the self-bias exhibited by the review models under the aforementioned different paradigms, this embodiment proposes the Relative Assessment Bias Coefficient (RABC). This index measures the systematic leniency tendency shown by a mental illness review model when evaluating its own generated content compared to evaluating content generated by other models. RABC is based on the ratio or difference between the false pass rate of reviewers for outputs from homologous models and the average false pass rate for outputs from heterologous models. Let... This indicates the rate at which reviewers overestimate the generator, and is defined as follows: (1).

[0038] in, This represents the set of samples to be evaluated. Represents a single sample. This indicates the specific criterion item for the corresponding sample. Indicates the reviewer For the sample In the The judgment result for each criterion item is 1, where 1 indicates "pass" and 0 indicates "failure"; This represents the objective reference value or programmed truth value of the corresponding item.

[0039] thus, Essentially, this means that among all the criteria that should be objectively judged as failures, the reviewer... However, the percentage that was incorrectly judged as passing, i.e., the false pass rate, can also be understood as a misjudgment of the generator's pass rate. The excessively loose rate is expressed as .

[0040] Based on this, this embodiment further defines the reviewer. Self-preference bias coefficient for: (2).

[0041] in, This represents a set of heterogeneous generative models other than the model family corresponding to the reviewer. This represents the cardinality of the set. If... Explain the reviewers When evaluating its own generated content, it exhibits a higher tendency to falsely approve compared to evaluating content generated by other models, indicating a self-biased bias; the larger the value, the stronger the bias. Conversely, if the value is close to 1, it indicates that the reviewer is generally neutral when faced with its own and other models' outputs; if it is less than 1, it indicates that the reviewer is actually more stringent about its own generated results.

[0042] S2.2: Using the self-preference bias coefficient based on fine-grained mental illness assessment criteria, statistical analysis is performed on the judgment results of the review model on candidate outputs, and the false pass rate, criterion tolerance, and local bias intensity are calculated on failed samples.

[0043] Among them, the criterion tolerance and the local bias intensity are both expressed by the false pass rate in equation (1) and the RABC in equation (2): the former reflects the difference in false pass between the output of the same source model and the output of the different source model on a single criterion, and the latter reflects the amplification of the difference relative to the average level of the different source.

[0044] In one specific implementation, at the criterion level, based on fine-grained mental illness assessment criteria, statistical analysis is performed on the judgment results of the review model on candidate outputs. The focus is on examining the distribution of reviewers' approval judgments for their own generated results or those generated by homologous models when the objective criteria conditions are not met. Based on this, the false approval rate, criterion tolerance, and local bias intensity on failed samples are calculated, thereby characterizing the implicit bias behavior of the model in specific medical safety criteria, empathy support criteria, risk identification criteria, or mental health ethics criteria.

[0045] This step calculates the false pass rate at the criterion level, the relative evaluation bias coefficient (RABC), and the instance-level consistency index based on the structured review results to quantify the reviewers' self-bias in their assessment of the outputs of the same-source model and the outputs of the different-source model. The false pass rate is the proportion of criterion items that were judged as failed by the reference results but passed by the review results, out of the total number of criterion items judged as failed by the reference results.

[0046] To enhance the robustness of this indicator in practical deployment, this embodiment may also introduce smoothing terms, confidence interval estimation, or hierarchical weighted statistical strategies in preferred implementations. For example, for criteria with a small number of failed samples, Bayesian smoothing can be used to avoid drastic fluctuations in false pass rates due to sample sparsity. For criteria with different severity levels, different mental illness categories, different risk levels, or different ethical sensitivities, different statistical weights can be set to improve the interpretability and stability of RABC in complex mental health scenarios. Especially when dealing with high-risk samples such as self-harm and suicide risks, misjudgments of psychotic symptoms, severe functional impairment, and psychological crises in minors, the system can increase the weights of relevant criteria to more accurately capture the model's self-biased performance on key medical safety issues.

[0047] Even in tasks with high objectivity and procedural verifiability, such as following instructions related to mental illness, ensuring the completeness of risk warnings, and adhering to medical safety regulations, self-bias is still prevalent. For example, the false pass rate of a mental illness review model when evaluating its own output is often significantly higher than its average false pass rate when evaluating the outputs of other models. This phenomenon indicates that the so-called "more lenient with its own conclusions" review bias is not only present in open-ended subjective tasks, but may be deeply embedded in the generation and discrimination coupling mechanism of mental illness models, the distribution preference of training corpora, the organization of psychiatric knowledge, and the internal representation structure. Therefore, explicit modeling and quantification of this type of bias is a prerequisite for improving the objectivity and reliability of automatic evaluation in large-scale mental illness models.

[0048] S2.3: Compare the order pair consistency between the output of the review model and the benchmark truth value to form a deviation measure index that reflects the global preference premium.

[0049] In one specific implementation, in addition to the deviation statistics at the criterion level mentioned above, this embodiment may also introduce indicators such as Mean Instance Pairwise Accuracy (MIPA) at the instance level to represent the consistency ratio between the candidate output ranking or quality judgment given by the review model and the benchmark true value.

[0050] Specifically, for multiple candidate answers to the same mental illness counseling input sample, the system compares the merits / dismerits given by the review model with the reference truth ranking pairwise to calculate the order pair consistency rate. The MIPA index can reflect the reliability of the review model at a more macro-level level of mental illness answer identification and can complement RABC: the former measures ranking accuracy, while the latter measures the magnitude of self-bias; the combination of the two can form a more comprehensive bias representation system.

[0051] This embodiment further compares the order-pair consistency between the review model output and the baseline truth value at the instance-level. By analyzing the reviewers' ranking relationships, judgments of merit, or scoring results of candidate answers from different sources and their consistency with the reference standard, it derives the overall deviation magnitude exhibited by the reviewers for specific model families, specific psychiatric expression styles, or specific source texts, forming a bias measure index reflecting the global preference premium. Through the above-mentioned two-level quantification mechanism, this embodiment can simultaneously achieve the joint representation of local criterion bias and overall instance bias.

[0052] S3: Based on the fine-grained evaluation criteria attributes, perform structured analysis and risk identification on the deviation inducing conditions to obtain the risk identification results of the criteria attributes.

[0053] S3.1: Extract and record various feature information of each fine-grained evaluation criterion. Specifically, perform attribute parsing on the fine-grained evaluation criteria, set structured analysis intervals for criterion lengths, automatically identify the semantic polarity of the criteria, and classify the criterion topics into domains.

[0054] In one specific implementation, this embodiment establishes a multidimensional correlation analysis mechanism to describe the relationship between the attributes of mental illness assessment criteria and the deviation coefficient. During the assessment process, various characteristic information of each criterion is automatically extracted and recorded, including but not limited to criterion text length, expression complexity, semantic positive or negative polarity, judgment direction, task category, risk sensitivity, professional domain affiliation, constraint strength, ethical sensitivity, subjective interpretation space, and whether it involves high-risk themes such as self-harm / suicide, violence risk, psychotic symptoms, and medication use recommendations. These characteristics are then jointly modeled with the improved relative evaluation deviation coefficient (RABC) to construct a correlation matrix between criterion attributes and deviation intensity between deviation factors and deviation responses. Based on this correlation matrix, high-risk criterion feature combinations that are more likely to induce model deviation in different mental health task scenarios are identified. This mechanism not only reveals the triggering conditions and distribution patterns of deviation but also provides a basis for automatic screening, risk labeling, structural optimization, and dynamic cleaning of the criterion library, reducing the probability of deviation from its source. From multiple perspectives, including criterion structure, semantic expression, task themes, and risk attributes, factors that may trigger self-bias are identified, labeled, and managed, thereby providing a basis for subsequent deviation detection, assessment optimization, and risk warning.

[0055] Specifically, the following steps are included: S3.1.1: First, perform attribute parsing on the fine-grained criteria.

[0056] Attribute analysis may include, but is not limited to, the following: guideline text length, syntactic complexity, semantic polarity, degree of logical nesting, judgment direction, task category, topic sensitivity, professional domain affiliation, constraint strength, medical risk level, and the size of the subjective interpretation space. Specifically, guideline text length reflects the scale of a single guideline; syntactic complexity indicates whether there are parallel conditions, progressive conditions, negative structures, adversative structures, or multiple exception clauses within the guideline; semantic polarity distinguishes whether the guideline is expressed as a positive requirement or a negative restriction; and topic sensitivity identifies whether the guideline involves high-risk topics such as mental illness risk identification, crisis intervention, medication recommendations, ethical boundaries, adolescent mental health, and self-harm / suicide warnings.

[0057] S3.1.2: Set the structured analysis interval for the criterion length.

[0058] For criteria that are too short, they are identified as potential candidate risk items with unclear semantic boundaries, insufficient targeting, or inadequate judgment criteria; for criteria that are too long, they are identified as potential candidate risk items with excessively long logical chains, easily overlooked local conditions, or increased model interpretation burden. Criteria can be further stratified and labeled according to a preset length threshold range, and then combined with RABC values, false pass rates, and review disagreement levels for joint analysis in subsequent deviation detection.

[0059] S3.1.3: Automatically identify the semantic polarity of the criteria.

[0060] Criteria expressed in terms such as "should include," "should be explicitly stated," or "requires prompting" can be classified as positive criteria; criteria expressed in terms such as "must not appear," "must not be downplayed," or "failure if not prompted" can be classified as negative criteria. Furthermore, secondary annotations can be added to the prohibited structures, negation conditions, defect identification requirements, and risk exposure requirements within negative criteria to identify the impact of specific types of criteria on the judgment behavior of the review model during deviation detection.

[0061] S3.1.4: Classify the criteria topics by domain.

[0062] In the context of mental illness assessment, the subject matter of guidelines is of high importance. Subject categories may include, but are not limited to: symptom identification, risk warning, crisis management, ethical compliance, medication boundaries, diagnostic caution, emotional support, patient communication, referral recommendations, and accuracy of medical facts. Guidelines involving self-harm, suicide, psychotic symptoms, acute agitation, severe insomnia, diminished reality testing ability, crises in minors, and perinatal mental disorders can be further designated as high-sensitivity mental health subject guidelines. In subsequent deviation management, these guidelines can be given higher monitoring priority, stricter risk thresholds, or more frequent review requirements.

[0063] S3.2: Construct a correlation matrix between the criterion attributes and the deviation intensity between the deviation factor and the deviation response based on the feature information, and obtain the deviation risk level, interpretation space level and suggested handling strategy corresponding to the fine-grained evaluation criteria.

[0064] In one specific implementation, to achieve a systematic correlation between the aforementioned attributes and deviation levels, this embodiment constructs a correlation matrix between criterion attributes and deviation intensity. This matrix maps criterion length, syntactic complexity, semantic polarity, topic category, risk sensitivity, subjectivity, judgment direction, and professional domain attributes to deviation response variables. Deviation response variables include, but are not limited to, the relative evaluation deviation coefficient (RABC), false pass rate, review consistency score, committee disagreement, outlier offset value, and re-evaluation trigger frequency. Regression analysis, kernel function mapping, tree models, graph structure modeling, cluster analysis, or rule engines can be used to model and update the relationship between criterion attributes and deviation responses.

[0065] Specifically, the steps for modeling and updating the relationship between criterion attributes and deviation responses are as follows: For each new review record, first determine the corresponding criterion attribute value and actual deviation response value.

[0066] Then update the matrix units according to the set learning rate. The update formula is: .

[0067] The learning rate can be preset based on the sample size, manual review results, or risk level of the criteria.

[0068] Specifically, for newly introduced mental illness assessment criteria, the system first extracts criterion attributes, then inputs the extraction results into a correlation matrix between criterion attributes and bias strength to obtain the corresponding bias risk level, explanatory space level, and suggested handling strategy. If the system identifies a criterion as possessing multiple risk attributes such as "abnormal length," "negative polarity," "highly sensitive topic," and "high subjectivity," it marks the criterion as a bias-sensitive criterion. In this case, the system can automatically generate optimization suggestions, including but not limited to: splitting the criterion, shortening sentences, reducing nested negations, introducing explicit triggering conditions, adding procedural verification items, setting manual review markers, or reducing its automatic judgment weight.

[0069] In this embodiment, the specific steps for extracting criteria attributes include: counting the length of the criteria text, identifying parallel conditions, negative structures and condition trigger words, determining the semantic polarity of the criteria, extracting mental health risk topics such as self-harm and suicide, drug safety, and crisis intervention, and marking the size of the subjective interpretation space.

[0070] When obtaining a strategy through the correlation matrix, the system reads the value of the criterion on each attribute item and its corresponding deviation response weight. If the weights of RABC or false pass rate corresponding to high-risk topics, negative polarity, and high subjectivity are high, the system outputs manual review or reduces the weight of automatic judgment. If the weights of abnormal length or nested negation are high, the system outputs splitting criteria, shortens sentence structure, or reduces nested negation. If the weight of committee disagreement is high, the system outputs secondary review or adds a heterogeneous review model.

[0071] To adapt the criterion attribute analysis mechanism to the dynamic changes in the large-scale mental illness model assessment, this embodiment also allows the system to be continuously updated based on historical review records. As new samples, criteria, and models are continuously added, the system can periodically adjust the risk coefficients, trigger thresholds, and early warning rules for various attributes, ensuring that the criterion attribute-bias strength correlation matrix remains scalable, transferable, and updatable. Through this mechanism, the assessment system can pre-identify and preprocess high-risk criteria before deviations actually occur, providing prior evidence for subsequent multi-model cross-validation and result correction modules.

[0072] S4: Based on the multi-model heterogeneous consensus detection mechanism, the bias of the review results is identified, isolated and corrected to obtain a consistency score.

[0073] In one specific implementation, to address potential issues such as broad homology bias, familial bias, expression consistency bias, and structural misjudgment in the actual review process of large-scale models of single mental illnesses, this embodiment designs a heterogeneous model consensus detection mechanism based on dynamic weight allocation. This mechanism introduces multiple review models from different technical approaches, training corpus sources, alignment methods, and parameter scales to form a review committee. Through parallel judgment and consistency analysis, it identifies bias risks in individual models. Specifically, based on the homology relationship between the review model and the candidate output generation model, model family relationships, historical bias records, or current abnormal review deviations, the corresponding review model is weighted down or eliminated through voting. More specifically, this embodiment dynamically allocates review weights to each model by comprehensively considering the historical stability, bias level, consistency with the human ground truth, and adaptability to specific mental illness tasks. The final review conclusion is formed using methods such as majority voting, weighted voting, confidence aggregation, or consistency optimization. Compared to relying on a single model for judgment, this mechanism can leverage the differences between heterogeneous models to offset individual biases and reduce the dominant influence of a particular family model on the mental illness assessment results.

[0074] The review committee comprises three or more heterogeneous review units. Each review unit can be a specialized model adapted from different base models for the field of mental illness, or a combination of general large models, medical large models, and mental health specialty models. Each review unit performs independent review for the same input sample, the same candidate output, and the same fine-grained criterion, outputting the corresponding binary judgment result, confidence score, explanation, or intermediate interpretation information. The review units are isolated from each other during the review phase to reduce the impact of crosstalk on the final consistency analysis.

[0075] For each specific criterion item, the outputs of all review units are first summarized, and the committee's consensus score for that item is calculated. In one feasible approach, this consensus score... It can be defined as: (3).

[0076] in, Indicates that each review unit is in the first... The voting results for each criterion item This indicates the majority opinion on this item. This indicates the total number of models participating in the review. The higher the value, the more unified the opinions of the review committee on that item, and the more stable the conclusion; conversely, if... A lower value indicates that there is significant controversy surrounding this item, and the judgment result may be affected by the ambiguity of the criteria, the complexity of the task, or model bias.

[0077] when If the value exceeds a preset threshold, such as 0.8, the system can directly adopt the committee consensus as the final conclusion for that criterion item; when... When the threshold is between a lower and a higher threshold, the system can perform weighted voting based on the historical performance of each model; while when... If the value falls below a set lower limit, such as 0.6, it can trigger manual intervention, multi-step reasoning reassessment, secondary review, or a high-reliability model verification mechanism to prevent disputed items from directly entering the final report.

[0078] To more effectively identify self-bias, this embodiment incorporates a bias stripping unit within the consensus detection mechanism. This bias stripping unit analyzes whether the current dispute is related to a particular review model and candidate outputs in terms of homology, model family relationships, similarity of training corpora, similarity of expression styles, or consistency of inference paths.

[0079] Specifically, the bias stripping unit first calculates the correlation between candidate outputs and each review model based on information such as model source identification, model family relationship, training corpus similarity, output expression style similarity, and inference path consistency. Then, it generates corresponding bias risk scores by combining the historical relative evaluation bias coefficient (RABC) records of each review model, the current criterion item judgment deviation magnitude, the degree of difference from the conclusions of other review models, the abnormally loose distribution on high-risk mental health criteria, and long-term review stability.

[0080] The judgment criteria may include: when the correlation between a certain review model and the candidate output is higher than a preset threshold, and the review model exhibits an abnormally lenient tendency or an abnormally consistent tendency that is significantly different from other review models on the current controversial item, the system determines that the review model has a potential self-biased influence on the candidate output. For the identified review models with bias risks, the system can perform processing such as weight suppression, vote reduction, anomaly marking, local elimination, delayed adoption, or triggering manual review to reduce their interference with the final review conclusion.

[0081] In some implementations, the deviation risk score can be obtained by weighted fusion of correlation score, historical deviation score, and current abnormal offset score, as shown below: (4).

[0082] in, This represents the correlation score between the candidate output and the review model. This represents the historical bias score obtained based on historical RABC records. This indicates the deviation score of the current review result relative to the overall conclusion of the committee. , , These are preset weighting coefficients.

[0083] The bias stripping unit can also represent the relationship between the review model and the generated model using a graph structure. The system constructs a correlation graph by including the review model, the generated model, the model family, the training source, and the domain adaptation label, and calculates the degree of homology between any review unit and the current candidate output. When this correlation exceeds a preset threshold, the system can automatically increase the bias review level of the review result and include it in the key monitoring scope. In this way, the system can identify potential bias paths before the review conclusion is formed, thereby avoiding an overly positive judgment on the candidate output by a model due to "similar source" or "similar style".

[0084] In addition to integrating review results, this invention also establishes a consistency-driven noise filtering process. For assessment items that consistently exhibit high discrepancies, high fluctuations, repeated reviews, and overlap with high-risk mental health criteria attributes, the system can mark them as bias-sensitive items or noise pollution items. For such items, one or more of the following processing strategies can be adopted: increasing the number of heterogeneous review models, introducing additional procedural validators, calling structured medical rule bases, performing chained interpretation reviews, reducing their weight in the overall score, setting manual review markers, or suspending the purely automatic assessment process for this item and transferring it to a manual handling process. The above processing methods can be configured differently according to the risk level of the mental illness task.

[0085] This embodiment introduces a consistency-driven noise filtering strategy to automatically identify and mark abnormal evaluation items in the multi-model review results that have significant disagreements, poor stability, or are suspected to be driven by self-bias, thereby improving the objectivity, robustness, and statistical reliability of the final evaluation report.

[0086] Multi-model consensus detection is used not only to form the final review conclusions but also to retrospectively track the bias behavior of individual models. Based on the deviation between the committee's overall judgment and the judgment of individual review models, it can identify the bias patterns of individual models under different criteria, task categories, and mental illness themes, and update the corresponding records in the historical bias archive. The historical bias archive can be further used for subsequent dynamic weight adjustments, risk criterion adaptation, model admission screening, and review credibility management, enabling the entire large-scale mental illness model evaluation system to possess continuous learning and closed-loop optimization capabilities.

[0087] S5: Based on the deviation quantification results, the criterion attribute risk identification results, and the consistency score, the review results are corrected for deviations and the final deviation detection results are obtained.

[0088] In one specific implementation, after completing deviation quantification, criterion attribute analysis, and multi-model consensus testing, this embodiment provides a correction and output control mechanism for the review results of a large-scale mental illness model. This mechanism integrates the deviation information, risk information, consistency information, and structured review results output from the aforementioned modules. For assessment items with high discrepancies, high volatility, and overlap with high-risk mental health criterion attributes, it performs re-evaluation, procedural verification, manual review, score downgrading, or transfer to manual review. This results in a more robust, interpretable, and applicable final assessment conclusion suitable for high-risk medical scenarios.

[0089] Specifically, the system first receives RABC values, false pass rates, instance-level consistency indicators from the deviation quantification modeling module, and risk classification results from the criterion attribute correlation analysis module. Then, it combines these with the consistency score, divergence markers, and deviation stripping information output by the multi-model consensus detection module to establish a comprehensive credibility score for each criterion item. The comprehensive credibility score characterizes the acceptability, controversy level, and whether subsequent human intervention is required for the criterion item's evaluation results. Based on the comprehensive credibility score, evaluation items can be processed in a tiered manner. For criterion items with high credibility, low controversy, and low deviation risk, their automatic evaluation results can be directly retained. For criterion items with medium credibility, secondary weighted correction or a prompt for re-evaluation can be performed. For criterion items with low credibility, high deviation risk, or high controversy, procedural rule review, human expert review, or delayed output mechanisms can be triggered. Especially for high-risk criteria involving self-harm / suicide, psychotic symptoms, drug safety, and acute crisis management, even after automatic evaluation results have been generated, additional safety thresholds and interception mechanisms can be set to ensure that the output results do not directly enter the downstream decision-making process due to model bias.

[0090] In this embodiment, the overall credibility score is obtained by subtracting the risk penalty factor from the positive credibility factor; wherein, the consistency score of MIPA and Equation (3) is used as the positive credibility factor, and RABC, false pass rate, criterion risk level, committee disagreement marker and deviation stripping information are used as risk penalty factors.

[0091] Results from different sources are first standardized to a range of 0 to 1: pass / fail judgments are directly converted to binary values, scores and consistency indicators are normalized according to maximum and minimum values, and high-risk entities and negative structures in the textual reasons are converted into risk labels. Subsequently, a comprehensive credibility score is obtained by weighting and summing the results according to preset weights; for high-risk criteria such as self-harm / suicide, psychotic symptoms, and drug safety, the weight of the criterion risk level and bias stripping information is increased.

[0092] When the overall credibility score is high and the risk of deviation is low, the automatic evaluation result is retained; when the overall credibility score is in the middle, a second weighted correction or re-evaluation is performed; when the overall credibility score is low, the criterion risk level is high, or the deviation stripping information is abnormal, manual expert review or delayed output is triggered.

[0093] The system can also generate structured output reports from the evaluation results. These reports include at least: sample identifiers, candidate output sources, judgment results for each criterion item, corresponding truth value comparison results, RABC correlation values, committee consensus scores, criterion risk labels, bias sensitivity markers, and final correction conclusions. For review items identified as having a risk of self-bias, the system can attach risk warning information to the report for further analysis by developers, reviewers, or mental health professionals. This structured output approach enhances the transparency, traceability, and auditability of the large-scale mental illness model evaluation process.

[0094] This embodiment allows for access control and usage restrictions on the final output. In high-risk mental health application environments, the system can restrict low-reliability assessment results to "for reference only," "requires manual review," or "cannot be directly used for clinical decision-making." For assessment results that pass bias detection and have high consistency, they can proceed to model training feedback, version comparison, data screening, or benchmark analysis processes. In this way, the present invention not only achieves self-bias detection in the fine-grained criteria assessment of large-scale mental illness models but also establishes an output control mechanism adapted to medical safety requirements.

[0095] Example 2: Embodiment 2 of the present invention provides a bias detection system for large-scale assessment of mental illness, comprising: The fine-grained criterion review module is used to call at least one review model to review multiple candidate outputs corresponding to the input sample under each fine-grained criterion item, and obtain structured review results.

[0096] The deviation quantification modeling module is used to obtain the fine-grained evaluation criteria and review results of the large model of the mental illness to be tested. It uses the self-preference deviation coefficient to quantify the deviation of the review results based on the output of the same source model and the output of the different source model according to the fine-grained evaluation criteria, and obtains the deviation quantification result.

[0097] The deviation quantification modeling module also includes: Construct a self-preference bias coefficient based on the relative evaluation bias coefficient index; Using the self-preference bias coefficient based on fine-grained mental illness assessment criteria, we conducted statistical analysis on the judgment results of the review model on candidate outputs, and calculated the false pass rate, criterion tolerance, and local bias intensity on failed samples.

[0098] The order pair consistency between the output of the review model and the benchmark truth is compared to form a deviation measure index that reflects the global preference premium.

[0099] The criterion attribute correlation analysis module is used to perform structured analysis and risk identification of deviation inducing conditions based on fine-grained evaluation criterion attributes, and obtain the criterion attribute risk identification results.

[0100] The criterion attribute correlation analysis module also includes: Extract and record various feature information of each fine-grained evaluation criterion. Specifically, perform attribute parsing on the fine-grained evaluation criterion, set structured analysis intervals for criterion length, automatically identify the semantic polarity of the criterion, and classify the criterion topics by domain. Based on the feature information, a correlation matrix between the criterion attributes and the deviation intensity between the deviation factor and the deviation response is constructed to obtain the deviation risk level, interpretation space level and recommended handling strategy corresponding to the fine-grained evaluation criteria.

[0101] The multi-model consensus detection module is used to identify, isolate, and correct deviations in the review results based on the multi-model heterogeneous consensus detection mechanism, and obtain a consistency score.

[0102] The multi-model consensus detection module also includes: Multiple review models from different technical approaches, training corpus sources, alignment methods, and parameter scales are introduced to form a review committee. Through parallel judgment and consistency analysis, the bias risks in a single model are identified. Specifically, based on the homology between the review model and the candidate output generation model, model family relationship, historical bias records, or current abnormal bias in the review, the corresponding review model is subjected to weight reduction or voting elimination.

[0103] The evaluation result correction output module is used to correct the deviation of the review results based on the deviation quantification results, the criterion attribute risk identification results, and the consistency score, and obtain the final deviation detection results.

[0104] The steps and methods involved in the above embodiment two correspond to those in embodiment one. For specific implementation details, please refer to the relevant description section of embodiment one.

[0105] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed in this application can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0106] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the flow or function according to the embodiments of this application is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in or transmitted through a computer-readable storage medium. The computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired or wireless means. The computer-readable storage medium can be any available medium that a computer can access or a data processing device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium, an optical medium, or a semiconductor medium, etc.

[0107] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for bias detection in the evaluation of a large model of mental illness, characterized in that, Includes the following steps: The fine-grained evaluation criteria and review results of the large model of the mental illness to be detected are obtained. The self-preference bias coefficient is used to quantify the bias of the review results based on the output of the same model and the output of the different models according to the fine-grained evaluation criteria, so as to obtain the bias quantification result. Based on the fine-grained evaluation criteria attributes, a structured analysis and risk identification of deviation inducing conditions are performed to obtain the risk identification results of the criteria attributes; Based on a multi-model heterogeneous consensus detection mechanism, biases in the review results are identified, isolated, and corrected to obtain a consistency score. Based on the deviation quantification results, the criterion attribute risk identification results, and the consistency score, the review results are corrected for deviations, and the final deviation detection results are obtained.

2. The bias detection method for the large-scale assessment of mental illness as described in claim 1, characterized in that, At least one review model is invoked to review multiple candidate outputs corresponding to the input sample under each fine-grained criterion, resulting in a structured review result.

3. The bias detection method for the large-scale mental illness model assessment as described in claim 1, characterized in that, The specific steps for quantifying the bias in the review results based on the outputs of the same-source model and the heterogeneous model using the self-preference bias coefficient and fine-grained evaluation criteria are as follows: Construct a self-preference bias coefficient based on the relative evaluation bias coefficient index; Using the self-preference bias coefficient based on fine-grained mental illness assessment criteria, statistical analysis was performed on the judgment results of the review model on candidate outputs, and the false pass rate, criterion tolerance, and local bias intensity were calculated on failed samples. The order pair consistency between the output of the review model and the benchmark truth is compared to form a deviation measure index that reflects the global preference premium.

4. The bias detection method for the large-scale mental illness model assessment as described in claim 1, characterized in that, The specific steps for performing structured analysis and risk identification of deviation-inducing conditions based on the fine-grained evaluation criterion attributes to obtain the risk identification results of the criterion attributes are as follows: Extract and record various feature information of each fine-grained evaluation criterion. Specifically, perform attribute parsing on the fine-grained evaluation criterion, set structured analysis intervals for criterion length, automatically identify the semantic polarity of the criterion, and classify the criterion topics by domain. Based on the feature information, a correlation matrix between the criterion attributes and the deviation intensity between the deviation factor and the deviation response is constructed to obtain the deviation risk level, interpretation space level and recommended handling strategy corresponding to the fine-grained evaluation criteria.

5. The bias detection method for the large-scale assessment of mental illness as described in claim 1, characterized in that, The specific steps for identifying, isolating, and correcting biases in the review results based on a multi-model heterogeneous consensus detection mechanism to obtain a consistency score are as follows: Multiple review models from different technical approaches, training corpus sources, alignment methods, and parameter scales are introduced to form a review committee. Through parallel judgment and consistency analysis, the bias risks in a single model are identified. Specifically, based on the homology between the review model and the candidate output generation model, model family relationship, historical bias records, or current abnormal bias in the review, the corresponding review model is subjected to weight reduction or voting elimination.

6. A bias detection system for evaluating a large-scale model of mental illness, characterized in that, include: The bias quantification modeling module is used to obtain the fine-grained evaluation criteria and review results of the large model of the mental illness to be detected. It uses the self-preference bias coefficient to quantify the bias of the review results based on the output of the same source model and the output of the different source models according to the fine-grained evaluation criteria, and obtains the bias quantification result. The criterion attribute correlation analysis module is used to perform structured analysis and risk identification of deviation inducing conditions based on fine-grained evaluation criterion attributes, and obtain the criterion attribute risk identification results. The multi-model consensus detection module is used to identify, isolate, and correct deviations in the review results based on the multi-model heterogeneous consensus detection mechanism, and obtain a consistency score. The evaluation result correction output module is used to correct the deviation of the review results based on the deviation quantification results, the criterion attribute risk identification results, and the consistency score, and obtain the final deviation detection results.

7. The bias detection system for the large-scale assessment of mental illness as described in claim 6, characterized in that, It also includes a fine-grained criterion review module, which calls at least one review model to review multiple candidate outputs corresponding to the input sample under each fine-grained criterion item, and obtains structured review results.

8. The bias detection system for the large-scale assessment of mental illness as described in claim 6, characterized in that, The deviation quantification modeling module also includes: Construct a self-preference bias coefficient based on the relative evaluation bias coefficient index; Using the self-preference bias coefficient based on fine-grained mental illness assessment criteria, statistical analysis was performed on the judgment results of the review model on candidate outputs, and the false pass rate, criterion tolerance, and local bias intensity were calculated on failed samples. The order pair consistency between the output of the review model and the benchmark truth is compared to form a deviation measure index that reflects the global preference premium.

9. The bias detection system for large-scale mental illness model assessment as described in claim 6, characterized in that, The criterion attribute correlation analysis module also includes: Extract and record various feature information of each fine-grained evaluation criterion. Specifically, perform attribute parsing on the fine-grained evaluation criterion, set structured analysis intervals for criterion length, automatically identify the semantic polarity of the criterion, and classify the criterion topics by domain. Based on the feature information, a correlation matrix between the criterion attributes and the deviation intensity between the deviation factor and the deviation response is constructed to obtain the deviation risk level, interpretation space level and recommended handling strategy corresponding to the fine-grained evaluation criteria.

10. The bias detection system for large-scale mental illness model assessment as described in claim 6, characterized in that, The multi-model consensus detection module also includes: Multiple review models from different technical approaches, training corpus sources, alignment methods, and parameter scales are introduced to form a review committee. Through parallel judgment and consistency analysis, the bias risks in a single model are identified. Specifically, based on the homology between the review model and the candidate output generation model, model family relationship, historical bias records, or current abnormal bias in the review, the corresponding review model is subjected to weight reduction or voting elimination.

Images