Explicit and implicit social bias propagation suppression method and system for multi-agent system

Abstract

The application discloses a kind of explicit and implicit social prejudice propagation inhibition method and system for multi-agent system.The method constructs three-dimensional prejudice island structure to depict the propagation path of prejudice in multi-round communication, extracts target agent prejudice dependency relationship through Bias Island Slice, and establishes a dynamic communication regulation mechanism based on Debias-Former.Further combined with reflection correction, viewpoint orthogonalization and identity anonymization, explicit and implicit prejudice joint inhibition is realized.The method does not need to change the original communication topology, and can be applied to dynamic multi-agent interaction environment, which can effectively reduce the propagation of social prejudice and improve task performance.

Description

Technical Field

[0001] The invention belongs to the field of artificial intelligence security and trusted computing technology, and in particular relates to methods and systems for suppressing the propagation of explicit and implicit social biases in multi-agent systems. Background Technology

[0002] In recent years, large language models (LMs) have demonstrated powerful capabilities in natural language understanding, reasoning and decision-making, and collaborative complex tasks. Multi-agent systems built on large language models achieve complex task solving, distributed reasoning, and group decision-making through the interaction and collaboration among multiple agents.

[0003] However, due to the prevalent social biases in pre-training corpora, multi-agent systems based on large language models inevitably inherit latent biases, which are further propagated and amplified during multiple rounds of communication. Existing bias propagation mainly manifests in two ways: explicit social bias propagation means that the biased content is directly reflected in the agent's output and propagates along the communication path. Implicit social bias propagation means that the bias is not directly expressed in the explicit output, but rather influences the subsequent agent's reasoning process through identity bias trust allocation, conformity effects, and majority opinion reinforcement.

[0004] Existing technologies have revealed several insurmountable limitations in addressing the critical need to suppress the propagation of two forms of social bias. First, existing methods lack dynamic propagation modeling capabilities; most only detect single-round outputs, failing to characterize the dynamic diffusion process of bias across time and topological dimensions. Second, they cannot jointly handle explicit and implicit biases, typically only addressing one type: content-level bias or structural-level bias. Finally, many methods rely on fixed communication structures, requiring modifications to the MAS topology, making them unsuitable for dynamic interaction systems. Therefore, a method is needed that can jointly detect and suppress dynamically propagating biases while maintaining the original multi-agent structure. Summary of the Invention

[0005] To address the aforementioned issues, this invention discloses a method and system for suppressing the propagation of explicit and implicit social bias in multi-agent systems. This method does not require changes to the original communication topology, is applicable to dynamic multi-agent interaction environments, and can effectively reduce the propagation of social bias and improve task performance.

[0006] The present invention provides a method for suppressing the propagation of explicit and implicit social biases in multi-agent systems, comprising the following steps:

[0007] Step 1: Interaction recording and collection, recording the communication behavior of all agents in each round.

[0008] Step 2: Spatiotemporal index mapping, mapping communication events to... Corresponding coordinates;

[0009] Three-dimensional space The X and Y axes represent the set of nodes of the agent, and the Z axis represents the discrete time step. Record the temporal evolution of interactions. Each node stores the time sequence from time step z to z+1, from the agent's... arrive Communication information:

[0010]

[0011] in respectively intelligent agents and Role types This represents the set of all possible roles in the system. It is essential for subsequent processing of conformity and identity discrimination within implicit biases. For intelligent agents The output content at this time step. Indicates in Does time exist from arrive The directed edge.

[0012] Step 3: Target node localization, determining the target intelligent agent. and the current round ;

[0013] Step 4: Recursive dependency backtracking, recursively backtracking along the topological connections to all its upstream dependent nodes.

[0014] Regarding time Target intelligent agent Its upstream interactions can be recursively extracted from the offset islands as a subgraph, which only spans the paths leading to the islands. The proxy nodes and time steps on the directed path:

[0015] ;

[0016] Step 5: Forming the initial bias island slice .

[0017] Step 6: For the time step Target Agent Construct the input sequence:

[0018] ,

[0019] And include a collection of perspectives on social biases in the discussion topics. To form a complete sequence:

[0020] .

[0021] Step 7: Perform undifferentiated discussion initialization training to learn basic bias identification capabilities.

[0022] Step 8: Perform differentiated discussion reinforcement training to establish a topology constraint communication mechanism.

[0023] A bias isolation mask is introduced in the Transformer attention layer to independently isolate the biased reference space. Within the biased view set, mutual attention is allowed, but ordinary communication nodes cannot interfere with the biased reference space in the reverse direction. The bias isolation mask is represented as follows:

[0024]

[0025] Island-aware masks are dynamically generated based on the bias island slices to restrict the scope of information flow, allowing only nodes that satisfy topological dependence on temporal causality to participate in attention computation. The island-aware mask is represented as:

[0026]

[0027] Wherein the interaction density operator: ;

[0028] Reinforcement learning is employed to optimize parameters, constructing three types of reward signals: format constraint reward, semantic correctness reward, and bias-aware reward. Network parameters are updated via policy gradient. The bias reward function is:

[0029] ;

[0030] in It is the maximum time step, and This function measures the maximum semantic similarity between viewpoints. It measures the bias in the output by calculating the similarity between the agent's output and biased thinking.

[0031] Step 9: Bias node prediction.

[0032] Step 10: Project to generate refined slices.

[0033] Step 11: Determine the earliest bias node.

[0034] Step 12: Perform explicit bias reflection and correction, and update the local output.

[0035] Step 13: Perform viewpoint orthogonalization and remove redundant majority edges.

[0036] For agents with the same role and the same viewpoint, perform viewpoint orthogonalization by taking their outputs:

[0037]

[0038] in Indicates the time step The set of intelligent agents holding the majority viewpoint. Indicates the time step A set of intelligent agents with the same identity.

[0039] Step 14: Perform identity anonymization and evenly distribute the propagation weights to a uniform value. Remove the sender role labels, including expert identity, age attribute, gender label, and historical credibility identifier, from all remaining communication edges, and redistribute the propagation weights evenly.

[0040] Step 15: Replay the downstream reasoning and update the propagation chain.

[0041] Step 16: Generate the final debiased output.

[0042] The explicit and implicit social bias propagation suppression system for multi-agent systems provided by this invention includes an explicit and implicit bias propagation modeling module, a bias identification module, and an explicit and implicit bias removal module, as detailed below:

[0043] The explicit and implicit bias propagation modeling module is used to collect interaction information during the operation of a multi-agent system in real time, map the communication behavior between agents into a dynamic graph topology, and map all communication events into a three-dimensional bias island structure.

[0044] This module further includes the following units:

[0045] The communication graph construction unit, deeply coupled with the multi-agent scheduling framework, sets up interactive listening interfaces in the agent message scheduling layer, inference invocation layer, and state synchronization layer to monitor message passing behavior between agents in real time, obtaining detailed information such as the sending agent identifier, receiving agent identifier, communication round, communication content, communication trigger time, and role attributes. When any agent completes its current inference round and sends information to other agents, the listening interface immediately captures the communication event and extracts information such as the sending node identifier, receiving node identifier, communication round, agent role, output content, and timestamp. The collected data is uniformly stored in the communication event buffer. The communication graph construction unit builds a directed dynamic graph structure based on the communication records in the buffer, where each agent instance is mapped to a graph node, each communication behavior is mapped to a directed edge, the edge direction represents the information propagation direction, and the edge weight represents the strength of the communication preference.

[0046] The Bias Island Generation Unit is used to perform a three-dimensional index mapping operation on all communication records and establish a Bias Island structure after the communication graph is constructed. During system execution, whenever a new communication event is detected, the Bias Island Generation Unit automatically determines the three-dimensional coordinate position based on the sending node, receiving node, and time cycle, and writes the communication content into the corresponding storage slot.

[0047] The Bias Island Slice Extraction Unit is used to recursively extract the local propagation dependency structure of the target agent from the Bias Island. After receiving the target agent identifier and the current round, this unit first locates the coordinates of the target node in the Bias Island. Then, starting from the target node, it backtracks layer by layer along all upstream edges that meet the connection conditions until the initial round is reached or no valid upstream node exists, thus forming the initial Bias Island slice.

[0048] The bias identification module is used to convert the bias island slices into a structured input sequence, introduce a set of social bias perspectives as a bias reference context, constrain the information flow range through attention masking, and output a refined bias island slice labeled with bias nodes.

[0049] This module further includes the following units:

[0050] The input sequence construction unit, upon receiving a bias slice, performs topological sorting of the communication nodes within the slice based on communication topology and temporal dependencies. The sorting principles include: prioritizing temporal causal order; ensuring upstream nodes take precedence over downstream nodes; and ensuring nodes in the same round are grouped by role. Subsequently, the communication content corresponding to each node is extracted, converted into a unified semantic embedding representation, and a set of social bias perspectives is introduced as a bias reference context. A structured sequence in a unified format is then output to the subsequent recognition network.

[0051] The bias isolation training unit is used to establish the basic bias recognition capability of Debias-Former. This unit performs undifferentiated discussion training in the initial training phase, introducing a bias isolation mask in the Transformer attention layer to independently isolate the biased reference space. This allows mutual attention within the biased viewpoint set, while preventing ordinary communication nodes from interfering with the biased reference space. This is achieved by writing a negative infinity mask value into the attention score matrix. During training, the model parameters are optimized using supervised learning. The training labels are derived from a manually labeled set of biased nodes, and the optimization objective is to minimize the classification error between predicted biased nodes and true biased nodes.

[0052] The Debias-Former unit is used to perform topological constraint bias recognition during the actual inference phase. After the input sequence enters the Transformer, this unit dynamically generates an island-aware mask based on bias island slices, restricting the scope of information flow and allowing only nodes that satisfy topological dependence on temporal causality to participate in attention computation. The system first calculates the interaction density of roles and then constructs the island-aware mask. This unit uses reinforcement learning to optimize the Debias-Former parameters, constructing three types of reward signals: format constraint reward, semantic correctness reward, and bias-aware reward, and updates the network parameters through policy gradients.

[0053] The explicit and implicit bias removal module is used to process biased nodes in chronological topological order, perform explicit bias reflection and correction, viewpoint orthogonalization and identity anonymization, and replay the corrected propagation chain in chronological order to drive downstream agents to re-infer and generate the final bias-removed output.

[0054] This module further includes the following units:

[0055] The explicit bias reflection unit is used to perform content correction on communication nodes with explicit biases. After receiving the set of biased nodes, this unit processes each biased node in chronological topological order. For any biased node, its local communication context is extracted and the reflection inference operator is invoked to guide the corresponding agent to re-examine its own output and regenerate the inference result by combining it with upstream evidence.

[0056] The viewpoint orthogonalization unit is used to suppress the propagation of implicit bias caused by the herd effect. This unit first counts the distribution of all agents' views in the current round and determines the majority viewpoint set. Then, it filters the communication edges that belong to the same role and express the same majority viewpoint, retaining only one representative edge and removing all other redundant edges.

[0057] The identity anonymization unit is used to eliminate role authority bias. After orthogonalizing viewpoints, this unit uniformly removes the sender's role labels, including expert identity, age attribute, gender label, and historical credibility identifier, from all remaining communication edges, while simultaneously redistributing uniform propagation weights.

[0058] The re-inference execution unit is used to perform recursive inference on the corrected bias island slice. This unit replays the communication propagation process in chronological order, first reading the output of the corrected node and propagating it to downstream nodes, triggering downstream nodes to re-infer, and finally updating the slice state. This process continues until the target node completes its final inference.

[0059] After the system of this invention is started, it will collect the communication behavior of the multi-agent system in each round in real time, construct a directed dynamic graph, and map the interaction information to a three-dimensional bias island structure. For the target agent, it will recursively backtrack the upstream dependent nodes along the topological connections, extract the local propagation subgraph, and form the initial bias island slice.

[0060] The initial slices are further transformed into structured input sequences, incorporating a set of social bias perspectives as a reference. Differentiation is then discussed using island-aware masks, outputting refined bias island slices labeled with bias nodes.

[0061] In the final debiasing stage, the system processes biased nodes in temporal topological order: reflexive correction is performed on explicit biases, updating local outputs; orthogonalization is performed on the viewpoint distribution, removing redundant majority edges; and identity anonymization is performed on communication edges, balancing propagation weights. Finally, downstream inference is replayed in sequence to generate the final debiased output.

[0062] Beneficial effects of the present invention

[0063] 1. Achieving joint suppression of explicit and implicit social biases: This invention is the first to explicitly divide the propagation of social biases in multi-agent systems into two paths: explicit propagation and implicit propagation. It also uses bias island slicing to unify the structured representation of the two propagation mechanisms. Based on this, it suppresses the content replication and reinforcement of explicit biases through reflection and correction, and suppresses the conformity effect and identity bias trust allocation in implicit biases through opinion orthogonalization and identity anonymization. This achieves joint detection and synergistic suppression of the two types of biases, overcoming the limitation of existing methods that can only deal with a single type of bias.

[0064] 2. No need to change the original multi-agent communication topology: This invention takes dynamic control at the communication layer as its core. Without modifying the original composition of the agents and the communication structure of the multi-agent system, it can effectively suppress the propagation of bias through the attention masking mechanism of Debias-Former and the debias optimization strategy. This avoids the potential damage to the system function caused by existing topology pruning or structural reconstruction methods. It is applicable to multi-agent systems in various fixed communication topologies or dynamic interaction environments.

[0065] 3. Effectively reduce bias propagation and improve task performance: Experimental results on multiple multi-agent interaction paradigms (multi-expert teams, group discussions, debates) and multiple bias evaluation benchmarks show that the proposed method can continuously reduce social bias scores while maintaining or improving task accuracy, achieving a good balance between bias suppression and task performance, and verifying the effectiveness of communication-level dynamic regulation as a system-level bias mitigation paradigm.

[0066] 4. Strong dynamic propagation modeling capability: This invention uses a three-dimensional bias island structure to characterize the temporal evolution and topological diffusion process of bias in multi-round communication. By recursively extracting the complete upstream bias dependency chain of the target agent through bias island slices, it realizes a systematic representation of the dynamic propagation path of bias, overcomes the shortcomings of existing methods that only perform static detection on single-round outputs, and provides structured support for bias tracking and source tracing.

[0067] 5. Precise bias identification based on role perception and topological constraints: The present invention Debias-Former introduces an island-aware mask and a role interaction density operator. By jointly modeling topological dependence, temporal causality and role distribution, it restricts information flow to only the real propagation path, thereby achieving precise localization of bias propagation signals. At the same time, it maintains a stable bias reference context through a bias isolation mask, further improving the accuracy and robustness of bias identification. Attached Figure Description

[0068] Figure 1 This is the system architecture diagram of the present invention.

[0069] Figure 2 This is a flowchart illustrating the implementation of the bias recognition module of this invention.

[0070] Figure 3 This is a flowchart of the bias removal algorithm of the present invention. Detailed Implementation

[0071] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. It should be noted that the terms "front," "rear," "left," "right," "up," and "down" used in the following description refer to directions in the accompanying drawings, and the terms "inner" and "outer" refer to directions toward or away from the geometric center of a specific component, respectively.

[0072] like Figure 1 As shown, the explicit and implicit social bias propagation suppression system for multi-agent systems provided by this invention includes three core modules: an explicit and implicit bias propagation modeling module, a bias identification module, and an explicit and implicit bias removal module.

[0073] The bias identification module comprises an input sequence construction unit, a bias isolation training unit, and a differential reasoning identification unit. The input sequence construction unit simultaneously provides data for the latter two units, while the output of the isolation training unit serves as auxiliary input for differential reasoning identification. This module ultimately outputs a set of biased nodes and a refined slice of biased islands, which are then passed to the explicit / implicit bias removal module.

[0074] The bias removal module consists of four sequentially executed units: an explicit bias reflection unit, an opinion orthogonalization unit, an identity anonymization unit, and a re-inference execution unit. The output of the previous unit serves as the input of the next unit, forming a bias removal processing pipeline.

[0075] The overall data flow can be summarized as follows: the original multi-agent interaction data is used by the modeling module to generate bias island slices, then the identification module marks the bias nodes, and finally the debiasing module completes the correction and outputs the debiased final result.

[0076] In specific implementation, such as Figure 3 As shown, after the system starts, it first connects to the target multi-agent system operating environment and deploys communication monitoring interfaces in the agent scheduling layer, message exchange layer, and inference invocation layer. The system monitors the interaction behavior between all agents in real time and automatically records communication events, including the sending node identifier, receiving node identifier, communication round, communication content, role attributes, and timestamp information.

[0077] The collected communication events are continuously written to the communication buffer. The explicit / implicit bias propagation modeling module constructs a directed dynamic graph based on the buffered data and synchronously maps all interactions to a 3D bias island structure. Each 3D coordinate point records the communication content, role attributes, and connection relationships for the corresponding round, thus forming a global propagation topology storage space for the entire multi-round communication process. .

[0078] When the system receives a reasoning request from the target agent, the bias island slice extraction unit performs recursive topological backtracking based on the target agent's ID and the current round. This ultimately generates the initial bias island slice corresponding to the target agent.

[0079] .

[0080] This slice fully depicts the historical influence chain during the formation of the target output.

[0081] like Figure 2 As shown, the bias identification module performs topological and temporal sorting on the Bias Island Slice, organizing all communication nodes into a unified structured input sequence according to causal dependencies. }

[0082] Simultaneously, a set of perspectives on social biases is introduced. As a biased reference context, it is embedded in the input sequence to form the complete input of Debias-Former and incorporate a set of social bias perspectives on the topic of discussion. To form a complete sequence .

[0083] In the actual reasoning process, the system will dynamically generate an island-aware mask based on the Bias Island Slice to constrain the information flow.

[0084] .

[0085] The interaction density operator The network ultimately outputs a set of biased nodes and generates a refined slice of biased islands. .

[0086] For detected explicit bias nodes, the system calls the reflective inference operator one by one in temporal topology order. Each agent re-examines the existing inference results in conjunction with the local communication context, corrects and replaces the biased outputs, and weakens the propagation of explicit bias at the source.

[0087] For implicit propagation paths, the system performs two types of structural optimizations: redundant communication edges that repeatedly express the majority of viewpoints under the same role are removed, and only representative viewpoint edges are retained. ,

[0088] This reduces the bandwagon reinforcement effect. By removing sender role tags and identity attributes, all retained edges are assigned uniform propagation weights, eliminating identity authority bias. The system will then use the corrected bias island slices... The subsequent propagation chain is replayed. Each downstream agent sequentially reads the updated input, re-executes the inference, and updates the propagation state layer by layer until the target agent completes the final inference. This completes one full explicit-implicit social bias propagation suppression process.

[0089] Specific experimental verification section:

[0090] [Datasets] The system was validated on two benchmark datasets for assessing social bias: the BBQ dataset and the CALM dataset. For the BBQ dataset, we selected three representative bias dimensions in the context of de-biasing: gender, race, and age. For the CALM dataset, we evaluated the gender and race bias dimensions in a question-answering task. All experimental data were randomly sampled and strictly balanced, with a 1:1 ratio of socially advantageous to disadvantaged groups. The overall performance of the model was measured by accuracy, while the degree of bias was quantified using the bias scores officially defined for the BBQ and CALM datasets. All final experimental evaluations were completed using only the dataset's native evaluation metrics and ground truth labels. To assess implicit social bias, this paper references existing research methods and assigns gender and race identities to each agent in a group discussion multi-agent system; each identity is uniformly and randomly sampled within each hierarchical level. Related experiments were conducted on the unbiased GPQA benchmark dataset, using the trust difference index and accuracy as evaluation criteria.

[0091] [Large Language Model Selection and Baseline Comparison Methods] This paper selects three representative large language models as the underlying foundation for multi-agent systems for experimental evaluation: Qwen3-8B, Qwen3-32B, and GPT-4. Three types of multi-agent architectures supporting both collaborative and adversarial reasoning paradigms are constructed to compare the experimental performance of various methods.

[0092] Existing bias removal methods fall into two categories: one optimizes for a single large language model, and the other is based on a fixed-architecture multi-agent system. The latter is not suitable for the dynamic interaction scenario presented in this paper. Since there is currently no bias removal solution for dynamic multi-agent scenarios, this paper constructs a comparative baseline as follows: For the single-model bias removal algorithm (DeCAP), this method is uniformly applied to each agent; for system-level comparison, representative algorithms that can suppress the propagation of harmful information are selected as the baseline, including StateDelta, MADERA, and Resilience.

[0093] [Experiment Execution Details] The maximum generation length of a single agent was set to 1024, and the sampling temperature parameter was 0.7. All experiments were deployed on a single server with a hardware configuration of one NVIDIA RTX PRO 6000 (96GB) graphics card; the system environment was Ubuntu 22.04, and software dependencies included Python 3.12, PyTorch 2.3.0, and CUDA 12.1.

[0094] [Experimental Results] As shown in Table 1, the system achieves competitive performance compared to existing methods under various interaction paradigms: its overall average accuracy reaches 93.0% on the BBQ dataset and 94.4% on the CALM dataset. This method can consistently achieve performance improvements under different interaction paradigms and different bias types, fully verifying the effectiveness of the dynamic control mechanism at the communication level in improving the decision-making quality of multi-agent systems.

[0095] Table 1: Accuracy performance of the system compared to several comparative methods in multi-agent system paradigms of multi-expert discussion, group discussion, and debate. The underlying multi-agent frameworks include Qwen3-8b, Qwen3-32b, and GPT-4. Datasets include BBQ and CALM. SDT, DC, RS, and MD represent StateDelta, DeCAP, Resilience, and MADERA, respectively.

[0096] Table 1

[0097]

[0098] More importantly, the system proposed in this paper can effectively achieve the core goal of mitigating social bias. Some existing methods may experience bias reversal or exacerbation under certain experimental conditions, while this system will never introduce additional bias and can simultaneously weaken both explicit and implicit biases.

[0099] Taking the experimental results on the BBQ dataset (Table 2) as an example: the bias score performance of the system compared with several contrasting methods in multi-agent system paradigms of multi-expert discussion, group discussion, and debate. The underlying multi-agent system base includes Qwen3-8b, Qwen3-32b, and GPT-4. The dataset includes BBQ. SDT, DC, RS, and MD represent StateDelta, DeCAP, Resilience, and MADERA, respectively.

[0100] Table 2

[0101]

[0102] The baseline average bias scores for the multi-expert collaborative framework, group discussion mechanism, and debate mechanism were 0.42, 0.45, and 0.51, respectively. After system optimization, the bias scores for the three baselines decreased to 0.25, 0.28, and 0.31, respectively. In the classification task with the highest degree of baseline bias, the proposed method achieved the largest absolute reduction in bias.

[0103] The system can also mitigate implicit social bias (Table 3).

[0104] Table 3: System performance in mitigating implicit bias propagation Table 3

[0105]

[0106] On the GPQA dataset, the model improves prediction accuracy while eliminating trust bias caused by identity differences in all experimental systems.

[0107] The technical means disclosed in this invention are not limited to those disclosed in the above embodiments, but also include technical solutions composed of any combination of the above technical features.

Claims

1. A method for suppressing the propagation of explicit and implicit social biases in multi-agent systems, characterized by: Specifically, the steps include the following: Step 1: Collect interaction information during the operation of the multi-agent system in real time, map the communication behavior between agents into a dynamic graph topology, and map all communication events into a three-dimensional bias island structure; Step 2: For the target agent, starting from the target node, recursively backtrack along the topological connections to all upstream dependent nodes, extract the local propagation subgraph, and form the initial bias island slice; Step 3: Convert the bias island slice into a structured input sequence, introduce a set of social bias perspectives as a bias reference context, constrain the information flow range in attention calculation through island-aware mask, and output a refined bias island slice labeled with bias nodes; Step 4: Process the biased nodes in chronological topological order, perform reflective correction on explicit biased nodes, and perform viewpoint orthogonalization and identity anonymization on communication edges; Step 5: Replay the corrected propagation chain in chronological order to drive downstream agents to re-infer and generate the bias-free final output.

2. The method for suppressing the propagation of explicit and implicit social biases in multi-agent systems according to claim 1, characterized in that: The specific implementation of mapping communication events to the three-dimensional bias island structure in step 1 includes the following steps: Step 11: Establish a three-dimensional space ,in These correspond to the agent indices of the sender and receiver in agent communication, respectively, and the discrete time steps. This elevates the space to three dimensions, making it usable for recording the temporal evolution of interactions; Step 12: Store each coordinate point from the agent arrive The communication information during time step z to z+1 is represented as follows: ； in, respectively intelligent agents and Role types It is the set of all possible roles in the system; For intelligent agents The output content at this time step; Indicates whether there exists from arrive The directed edge.

3. The method for suppressing the propagation of explicit and implicit social biases in multi-agent systems according to claim 1, characterized in that: The specific implementation of extracting the initial bias island slice in step 2 includes the following steps: Step 21: For time Target intelligent agent Let the agent be numbered Its upstream interactions can be recursively extracted from the offset islands as a subgraph, which only spans the paths leading to the islands. The proxy nodes and time steps on the directed path: 。 4. The method for suppressing the propagation of explicit and implicit social biases in multi-agent systems according to claim 1, characterized in that: Step 3 involves constraining the information flow range using attention masks and optimizing bias recognition accuracy through reinforcement learning. This is specifically achieved through the following steps: Step 31: Construct the complete input sequence: ； ； This represents the problem of extracting the current input of a multi-agent system. Prejudice in society and the corresponding thought processes; Taken from the island of prejudice, for intelligent agents exist All upstream information at any given time; Step 32: Dynamically generate an island-aware mask based on the bias island slices. The mask is used to constrain the information flow range in the attention calculation; the constraints include: (a) Temporal causal constraint: Any agent can only pay attention to its own and all the historical information of its predecessor agents, and remains invisible to the communication content of its successor agents; (b) Topological connectivity constraint: Any agent can only pay attention to nodes with which it has a direct directed communication edge, and attention associations are prohibited between two nodes with no communication connection; Define the interactive density operator for: ； in for Identity at any time The modulus of the set of nodes, This is a preset concurrent communication threshold; when the number of agents with the same role is less than... hour, Used as a scaling factor for attention weights; when the quantity reaches or exceeds hour, Fixed at 1; Step 33: Introduce a bias isolation mask into the Transformer attention layer and perform supervised fine-tuning training. The mask independently isolates the bias reference space, allowing mutual attention within the biased viewpoint set, while preventing ordinary communication nodes from interfering with the bias reference space in the reverse direction. The bias isolation mask is represented as follows: ； Step 34: Further introduce bias island slices to dynamically generate island-aware masks in the Transformer attention layer and perform reinforcement learning fine-tuning. In the complete bias island slice dynamic generation of island-aware masks, right The attention value can be expressed as: ； in This indicates taking the current time step. This indicates the role of the agent corresponding to the current content. The value is 1 if the condition is true, and 0 otherwise. Step 35: Construct three types of reward signals, including format constraint reward, semantic correctness reward, and bias-aware reward, and update the network parameters through policy gradient; for the input sequence... ,set up The set of valid formats includes: Multi-agent time series labels: ; Empty tags: ;; The format constraint reward then manifests as follows: ； The semantic correctness reward is manifested as follows: ； in This indicates that the F1 score will be calculated for the output. Used to retrieve the truth value of this part of the output; the biased reward function is: ； in, This represents the total number of output sequences participating in the calculation in the current batch. This represents the maximum time step covered by the current island of bias slice. Used to measure the current output Collection of biased perspectives The maximum semantic similarity between them.

5. The method for suppressing the propagation of explicit and implicit social biases in multi-agent systems according to claim 1, characterized in that: The specific implementation of viewpoint orthogonalization in step 4 includes the following steps: Step 41: Calculate the distribution of viewpoints among all agents in the current round and determine the majority viewpoint set; for communication edges belonging to the same role and expressing the same majority viewpoint, retain only one representative edge and remove all other redundant edges, represented as: ； in Indicates the time step The set of intelligent agents holding the majority viewpoint. Indicates the time step A set of intelligent agents with the same identity; Step 42: Remove the sender role labels, including expert identity, age attribute, gender label and historical credibility identifier, from all retained communication edges, and redistribute the propagation weights evenly.

6. A system for suppressing the propagation of explicit and implicit social bias in multi-agent systems, based on the method for suppressing the propagation of explicit and implicit social bias in multi-agent systems according to any one of claims 1-5, characterized in that, include: The explicit and implicit bias propagation modeling module is used to collect interaction information during the operation of a multi-agent system in real time, map the communication behavior between agents into a dynamic graph topology, and map all communication events to a three-dimensional bias island structure. The explicit and implicit bias propagation modeling module includes a communication graph construction unit, a bias island generation unit, and a bias island slice extraction unit. The bias island slice extraction unit is used to recursively backtrack all upstream dependent nodes along the topological connection starting from the target node, extract local propagation subgraphs, and form initial bias island slices. The bias identification module is used to convert the bias island slices into structured input sequences, introduce a set of social bias perspectives as a bias reference context, constrain the information flow range through attention masks, and output refined bias island slices labeled with bias nodes; the bias identification module includes an input sequence construction unit, a bias isolation training unit, and a differential reasoning identification unit; The explicit and implicit bias removal module is used to process biased nodes in chronological topological order, perform explicit bias reflection and correction, viewpoint orthogonalization and identity anonymization, and replay the corrected propagation chain in chronological order to drive downstream agents to re-infer and generate the final bias-removed output; the explicit and implicit bias removal module includes an explicit bias reflection unit, a viewpoint orthogonalization unit, an identity anonymization unit and a re-inference execution unit.

Images