arxiv: 2604.08963 · v2 · submitted 2026-04-10 · 💻 cs.MA · cs.AI

Recognition: no theorem link

Aligned Agents, Biased Swarm: Measuring Bias Amplification in Multi-Agent Systems

Keyu Li , Jin Gao , Dequan Wang

Authors on Pith no claims yet

Pith reviewed 2026-05-10 17:19 UTC · model grok-4.3

classification 💻 cs.MA cs.AI

keywords multi-agent systemsbias amplificationecho chambersAI ethicssystemic polarizationworkflow structurescomparative judgmentsprejudice accumulation

0 comments

The pith

Multi-agent systems amplify bias through structured workflows rather than diluting it.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper investigates whether multi-agent collaboration reduces bias in AI decisions or makes it worse through organized interactions. It challenges the idea that adding more agents naturally leads to fairer outcomes by testing the possibility that workflows create self-reinforcing loops. Using a benchmark that requires direct comparisons across demographic groups, the study tracks how small initial biases grow larger over repeated steps. The results matter because real-world uses of these systems in hiring, lending, or content moderation could produce consistently skewed results if the amplification pattern holds.

Core claim

Contrary to the assumption that multi-agent collaboration naturally dilutes bias, structured workflows act as echo chambers that amplify minor stochastic biases into systemic polarization. Analysis of bias cascades across topologies shows that architectural sophistication frequently exacerbates bias rather than mitigating it, with systemic amplification occurring even when isolated agents operate neutrally and a trigger vulnerability where purely objective context accelerates polarization.

What carries the argument

The echo chamber effect in MAS topologies and feedback loops, where iterative agent interactions accumulate and reinforce minor stochastic biases into consistent polarization.

If this is right

Bias can reach systemic levels even when every individual agent begins neutral.
Greater architectural complexity in workflows tends to increase bias accumulation.
Injecting objective context can trigger faster polarization through the identified vulnerability.
Basic topologies and feedback loops alone suffice to produce measurable systemic effects.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Mitigation efforts for multi-agent systems may need to redesign interaction structures instead of relying only on individual agent alignment.
Similar echo chamber dynamics could appear in other collaborative setups such as sequential reasoning chains or tool-calling agents.
Applying the benchmark to larger real-world tasks would test whether amplification scales with problem size or domain.

Load-bearing premise

The forced comparative judgments in the benchmark accurately isolate structural bias amplification without introducing their own measurement artifacts or selection effects.

What would settle it

Repeated runs of the benchmark across multiple topologies showing no net rise in bias metrics when starting from neutral agents would falsify the amplification claim.

Figures

Figures reproduced from arXiv: 2604.08963 by Dequan Wang, Jin Gao, Keyu Li.

**Figure 1.** Figure 1: Two Parallel, Transformative Trends Shaping the Current AI Landscape. Left: The rapid advancement of powerful single-agent tooling, such as Codex (OpenAI, 2026) and Claude Code (Anthropic, 2026b), which excel in complex coding and generic problem-solving. Right: The paradigm shift towards complex Multi-Agent Systems (MAS), like Agent Teams (Anthropic, 2026a) and Agent Swarms (Kimi, 2026), designed for coll… view at source ↗

**Figure 2.** Figure 2: Overview of Our Framework for Investigating Iterative Bias Amplification in LLMbased MAS. Top: We propose Discrim-Eval-Open, an open-ended benchmark with multi-attribute options, to measure how an initial preference is progressively amplified as its reasoning is passed through a sequential chain of agents. Bottom: We then systematically evaluate whether common MAS architectures—employing diverse personas,… view at source ↗

**Figure 3.** Figure 3: Demographic Distribution of Protagonist Profiles in Discrim-Eval-Open Dataset. The benchmark includes 210 unique profiles with a diverse spread of attributes. Left: Age distribution covers a wide spectrum from individuals in their 20s to over 100. Center: Gender distribution is perfectly balanced, with exactly 70 instances each for Male, Female, and Non-binary identities. Right: Race and ethnicity distribu… view at source ↗

**Figure 4.** Figure 4: Systematic Evaluation of MAS Architectures as Potential Mitigators of Iterative Bias Amplification. Left: We investigate agent specialization in linear chains, testing whether assigning diverse personas (e.g., Doctor, Lawyer) and functions (e.g., Analyst, Reflector) can introduce varied perspectives to curb the amplification effect. Right: We evaluate the role of communication structure by designing more c… view at source ↗

**Figure 5.** Figure 5: Empirical Results Showing MAS Specialization Fails to Mitigate Iterative Bias Amplification. The plots show the relative Gini coefficient across four sequential agent layers (L1-L4) for eight different LLMs. (a) A baseline chain with identical roles confirms consistent amplification. Testing the mitigation hypothesis, we find that introducing (b) diverse personas, (c) specialized functions, or (d) a mix of… view at source ↗

**Figure 6.** Figure 6: MAS Architectural Complexity Fails to Mitigate but Exacerbate Bias Amplification. These plots show that complex communication structures and increased system depth do not solve the core issue of iterative amplification. (a-c) Bias progressively amplifies across all tested four-layer topologies (Spindle, Parallel, and Fully-Connected). (d) Furthermore, increasing system depth by iterating a fully-connected … view at source ↗

**Figure 7.** Figure 7: A Neutral Trigger Can Initiate a Cascade of Bias Amplification, Revealing System Fragility. This qualitative example compares two scenarios using a sequential MAS of Gemini 3.1 Pro (DeepMind, 2026) agents. Top Path: Without external input, the well-aligned system maintains a balanced and fair probability distribution. Bottom Path: However, introducing a single, seemingly objective sentence acts as a trigge… view at source ↗

**Figure 8.** Figure 8: MAS Tendency toward Favoring Younger Individuals, Women, and Black Communities. Results are derived from the whole benchmark across 70 scenarios, in a four-layer sequential MAS composed of DeepSeek-V3. Conclusion This work challenges the optimistic hypothesis that complex MAS architectures can mitigate the bias amplification inherent in multi-step LLM interactions. Our empirical findings, derived from… view at source ↗

**Figure 9.** Figure 9: Impact of Historical Information on Bias Amplification. MAS is constructed by sequentially connecting four agents using either DeepSeek-V3 or DeepSeek-R1. In the left subfigure, each agent receives the accumulated viewpoints from all preceding agents, whereas in the right subfigure, each agent only receives the opinion of its immediate predecessor. Results show that bias is progressively amplified in both … view at source ↗

**Figure 10.** Figure 10: Impact of Mixed Personas and Functions on Bias Amplification in MAS Construction. A four-agent MAS is constructed with a hybrid configuration: Agent 1 (left) serves as a judger, Agent 2 (top) as a doctor, Agent 3 (bottom) as an engineer, and Agent 4 (right) as a summarizer. Different LLMs are used to instantiate the agents, and variance is employed as the metric to quantify bias. Results show a clear tren… view at source ↗

**Figure 11.** Figure 11: Effect of Spindle MAS Topology on Bias Amplification, Measured by Variance. Agents 1–7 represent Judger, Doctor, Engineer, Summarizer, Lawyer, Merchant, and Summarizer, respectively. Lighter colors indicate higher variance, corresponding to more extreme bias. Results across multiple MAS configurations using different LLMs show that bias is progressively amplified, particularly between key functional nodes… view at source ↗

**Figure 12.** Figure 12: Impact of Iteration Rounds on Bias Amplification in MAS. The MAS is constructed using the same LLM across all nodes, with a topology consisting of four sequentially connected fully-connected sub-units. Higher variance indicates more extreme bias. The dashed baseline denotes the output of the first node (Judger) in the first sub-unit, while the solid lines represent the outputs of the final Summarizer node… view at source ↗

read the original abstract

While Multi-Agent Systems (MAS) are increasingly deployed for complex workflows, their emergent properties-particularly the accumulation of bias-remain poorly understood. Because real-world MAS are too complex to analyze entirely, evaluating their ethical robustness requires first isolating their foundational mechanics. In this work, we conduct a baseline empirical study investigating how basic MAS topologies and feedback loops influence prejudice. Contrary to the assumption that multi-agent collaboration naturally dilutes bias, we hypothesize that structured workflows act as echo chambers, amplifying minor stochastic biases into systemic polarization. To evaluate this, we introduce Discrim-Eval-Open, an open-ended benchmark that bypasses individual model neutrality through forced comparative judgments across demographic groups. Analyzing bias cascades across various structures reveals that architectural sophistication frequently exacerbates bias rather than mitigating it. We observe systemic amplification even when isolated agents operate neutrally, and identify a 'Trigger Vulnerability' where injecting purely objective context drastically accelerates polarization. By stripping away advanced swarm complexity to study foundational dynamics, we establish a crucial baseline: structural complexity does not guarantee ethical robustness. Our code is available at https://github.com/weizhihao1/MAS-Bias.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper introduces Discrim-Eval-Open to track bias growth in multi-agent setups but the benchmark's forced comparisons may be creating the polarization it reports.

read the letter

The paper's main point is that multi-agent structures can amplify small biases into larger polarization instead of smoothing them out, backed by a new open benchmark called Discrim-Eval-Open that uses forced comparative judgments across demographic groups. It also flags a trigger effect where neutral context speeds up the shift and shows amplification even from single neutral agents across different topologies and loops. The code is public on GitHub, which is a clear plus for anyone who wants to rerun or adapt the tests. That release and the baseline hypothesis give the work a practical starting point in an area where most claims stay high-level. The observations on neutral-agent baselines and the trigger vulnerability are the parts that stand out as worth checking against other setups. The soft spot is the benchmark design itself. Forcing agents into direct comparisons between groups can push outputs toward extremes by the way the questions are framed, which risks mixing benchmark artifacts with the claimed echo-chamber mechanism from topologies or feedback. The abstract gives no sample sizes, variance numbers, or controls for prompt wording, so it is difficult to judge whether the reported amplification is robust or mainly test-induced. If the full text has those details they need to be shown clearly. This is the kind of paper that researchers working on AI ethics and multi-agent coordination would want to see in a reading group to discuss the benchmark and run their own checks on the code. A reader looking for empirical cautions on MAS deployment could extract value from the public materials even if the claims need tightening. It deserves a serious referee to examine the experimental controls and prompt templates. I would send it to peer review with a request to add quantitative checks on whether the forced-judgment format is driving the results.

Referee Report

1 major / 1 minor

Summary. The manuscript claims that structured multi-agent system (MAS) workflows and feedback loops function as echo chambers, amplifying minor stochastic biases into systemic polarization rather than diluting them through collaboration. To investigate this, the authors introduce Discrim-Eval-Open, an open benchmark relying on forced comparative judgments across demographic groups to evaluate bias in various MAS topologies. Key observations include bias amplification even when individual agents are neutral, greater exacerbation with increased architectural sophistication, and a 'Trigger Vulnerability' in which purely objective context accelerates polarization. The work positions itself as a foundational baseline study, with code released for reproducibility.

Significance. If the results are robust to benchmark artifacts, the paper would contribute meaningfully to AI ethics and MAS research by challenging the common assumption that multi-agent collaboration inherently reduces bias and by highlighting risks from structural complexity. The open benchmark, identification of trigger effects, and public code release are positive elements that could enable follow-up work. The significance is limited by the empirical nature of the study and the need to confirm that observed effects arise from MAS mechanics rather than benchmark design.

major comments (1)

The central claim that MAS topologies and feedback loops drive bias amplification rests on Discrim-Eval-Open accurately isolating structural effects. However, the benchmark's reliance on forced comparative judgments across demographic groups risks introducing its own selection pressure toward polarized outputs, which could explain the reported amplification in neutral-agent baselines and the 'Trigger Vulnerability' rather than the hypothesized echo-chamber mechanism. This design choice is load-bearing and requires either additional controls (e.g., non-forced judgment variants or alternative bias metrics) or explicit discussion of why artifacts are ruled out.

minor comments (1)

The abstract and early sections would benefit from explicit statements of sample sizes, number of independent runs, statistical tests used for bias quantification, and precise definitions of 'neutral' agent behavior to allow verification of the high-level findings.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for their thoughtful review and constructive comments on our work. Below, we address the major comment point by point, outlining the revisions we plan to make to strengthen the manuscript.

read point-by-point responses

Referee: The central claim that MAS topologies and feedback loops drive bias amplification rests on Discrim-Eval-Open accurately isolating structural effects. However, the benchmark's reliance on forced comparative judgments across demographic groups risks introducing its own selection pressure toward polarized outputs, which could explain the reported amplification in neutral-agent baselines and the 'Trigger Vulnerability' rather than the hypothesized echo-chamber mechanism. This design choice is load-bearing and requires either additional controls (e.g., non-forced judgment variants or alternative bias metrics) or explicit discussion of why artifacts are ruled out.

Authors: We thank the referee for highlighting this important methodological consideration. The forced comparative judgment format in Discrim-Eval-Open was deliberately selected to create a controlled setting where bias can be measured through explicit group comparisons, as open-ended tasks often result in agents defaulting to neutral or evasive responses even when underlying biases exist. This design draws from established practices in LLM bias evaluation benchmarks. Nevertheless, we agree that it is essential to address potential artifacts. In the revised version, we will add a dedicated subsection in the methodology (Section 3) discussing the rationale for this choice, including references to similar forced-choice paradigms in social psychology and AI fairness literature. We will also explicitly discuss the limitations and the possibility that the benchmark design contributes to observed effects, while arguing that the consistent patterns across different MAS topologies (e.g., greater amplification in more complex structures) provide evidence for structural influences beyond benchmark artifacts. We will not introduce new experimental controls in this revision due to the scope of the baseline study but will outline them as directions for future work. revision: partial

Circularity Check

0 steps flagged

No significant circularity: empirical benchmark study with independent observations

full rationale

The paper is a purely empirical investigation that introduces the Discrim-Eval-Open benchmark and reports experimental observations of bias amplification across MAS topologies. No equations, derivations, fitted parameters, or self-citations appear in the provided text, and the central claims rest on direct experimental outcomes rather than reducing to inputs by construction. The benchmark design and results are presented as independent measurements, satisfying the criteria for a self-contained empirical study with no load-bearing circular steps.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the assumption that the new benchmark validly measures structural bias effects and that observed polarization stems from topology rather than benchmark design or agent internals.

axioms (1)

domain assumption Forced comparative judgments across demographic groups in Discrim-Eval-Open bypass individual model neutrality and reveal true prejudice accumulation.
Invoked to justify the benchmark as isolating structural effects.

pith-pipeline@v0.9.0 · 5499 in / 1029 out tokens · 29593 ms · 2026-05-10T17:19:25.023568+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

28 extracted references · 23 canonical work pages · 9 internal anchors

[1]

Measuring implicit bias in explicitly unbiased large language models.arXiv preprint arXiv:2402.04105, 2024

Anthropic. Building a c compiler with a team of parallel claudes, 2026a. URL https://www. anthropic.com/engineering/building-c-compiler. Anthropic. Introducing claude opus 4.6, 2026b. URL https://www.anthropic.com/news/ claude-opus-4-6. Xuechunzi Bai, Angelina Wang, Ilia Sucholutsky, and Thomas L Griffiths. Measuring implicit bias in explicitly unbiased l...

work page arXiv
[2]

Towards implicit bias detection and mitigation in multi-agent llm interactions.arXiv preprint arXiv:2410.02584,

Angana Borah and Rada Mihalcea. Towards implicit bias detection and mitigation in multi-agent llm interactions.arXiv preprint arXiv:2410.02584,

work page arXiv
[3]

Jwala Dhamala, Tony Sun, Varun Kumar, Satyapriya Krishna, Yada Pruksachatkun, Kai-Wei Chang, and Rahul Gupta

URL https://blog.google/innovation-and-ai/models-and-research/ gemini-models/gemini-3-1-pro/. Jwala Dhamala, Tony Sun, Varun Kumar, Satyapriya Krishna, Yada Pruksachatkun, Kai-Wei Chang, and Rahul Gupta. Bold: Dataset and metrics for measuring biases in open-ended language genera- tion. InProceedings of the 2021 ACM conference on fairness, accountability,...

2021
[4]

Agentscope: A flexible yet ro- bust multi-agent platform.arXiv preprint arXiv:2402.14034,

Dawei Gao, Zitao Li, Xuchen Pan, Weirui Kuang, Zhijian Ma, Bingchen Qian, Fei Wei, Wenhao Zhang, Yuexiang Xie, Daoyuan Chen, et al. Agentscope: A flexible yet robust multi-agent platform. arXiv preprint arXiv:2402.14034,

work page arXiv
[5]

ChatGLM: A Family of Large Language Models from GLM-130B to GLM-4 All Tools

11 Published as a conference paper at ICLR 2026 Team GLM, Aohan Zeng, Bin Xu, Bowen Wang, Chenhui Zhang, Da Yin, Dan Zhang, Diego Rojas, Guanyu Feng, Hanlin Zhao, et al. Chatglm: A family of large language models from glm-130b to glm-4 all tools.arXiv preprint arXiv:2406.12793,

work page internal anchor Pith review arXiv 2026
[6]

DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning

Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948,

work page internal anchor Pith review Pith/arXiv arXiv
[7]

GPT-4o System Card

Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Os- trow, Akila Welihinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card.arXiv preprint arXiv:2410.21276,

work page internal anchor Pith review Pith/arXiv arXiv
[8]

Mapcoder: Multi-agent code generation for competitive problem solving.arXiv preprint arXiv:2405.11403,

Md Ashraful Islam, Mohammed Eunus Ali, and Md Rizwan Parvez. Mapcoder: Multi-agent code generation for competitive problem solving.arXiv preprint arXiv:2405.11403,

work page arXiv
[9]

Towards mitigating llm hallucination via self reflection

Ziwei Ji, Tiezheng Yu, Yan Xu, Nayeon Lee, Etsuko Ishii, and Pascale Fung. Towards mitigating llm hallucination via self reflection. InFindings of the Association for Computational Linguistics: EMNLP 2023, pp. 1827–1843,

2023
[10]

Harbor: exploring persona dynamics in multi-agent competition

Kenan Jiang, Li Xiong, and Fei Liu. Harbor: exploring persona dynamics in multi-agent competition. arXiv preprint arXiv:2502.12149,

work page arXiv
[11]

Keyu Li, Mohan Jiang, Dayuan Fu, Yunze Wu, Xiangkun Hu, Dequan Wang, and Pengfei Liu

URL https://www.kimi.com/blog/ kimi-k2-5. Keyu Li, Mohan Jiang, Dayuan Fu, Yunze Wu, Xiangkun Hu, Dequan Wang, and Pengfei Liu. Datasetresearch: Benchmarking agent systems for demand-driven dataset discovery.arXiv preprint arXiv:2508.06960,

work page arXiv
[12]

AgencyBench: Benchmarking the Frontiers of Autonomous Agents in 1M-Token Real-World Contexts

Keyu Li, Junhao Shi, Yang Xiao, Mohan Jiang, Jie Sun, Yunze Wu, Shijie Xia, Xiaojie Cai, Tianze Xu, Weiye Si, et al. Agencybench: Benchmarking the frontiers of autonomous agents in 1m-token real-world contexts.arXiv preprint arXiv:2601.11044,

work page internal anchor Pith review Pith/arXiv arXiv
[13]

DeepSeek-V3 Technical Report

Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, et al. Deepseek-v3 technical report.arXiv preprint arXiv:2412.19437, 2024a. Zhao Liu, Tian Xie, and Xueru Zhang. Evaluating and mitigating social bias for large language models in open-ended settings.arXiv preprint arXiv:2412.06134,...

work page internal anchor Pith review Pith/arXiv arXiv
[14]

BBQ: A hand-built bias benchmark for question answering.arXiv preprint arXiv:2110.08193,

URL https://github.com/openclaw/ openclaw. Alicia Parrish, Angelica Chen, Nikita Nangia, Vishakh Padmakumar, Jason Phang, Jana Thompson, Phu Mon Htut, and Samuel R Bowman. Bbq: A hand-built bias benchmark for question answering. arXiv preprint arXiv:2110.08193,

work page arXiv
[15]

Herding in humans.Trends in cognitive sciences, 13(10):420–428,

12 Published as a conference paper at ICLR 2026 Ramsey M Raafat, Nick Chater, and Chris Frith. Herding in humans.Trends in cognitive sciences, 13(10):420–428,

2026
[16]

Bias mitigation agent: Optimizing source selection for fair and balanced knowledge retrieval.arXiv preprint arXiv:2508.18724,

Karanbir Singh, Deepak Muppiri, and William Ngu. Bias mitigation agent: Optimizing source selection for fair and balanced knowledge retrieval.arXiv preprint arXiv:2508.18724,

work page arXiv
[17]

Multi-agent collaboration: Harnessing the power of intelligent llm agents.arXiv preprint arXiv:2306.03314, 2023

URLhttps://www.stepfun.com/#step1v. Yashar Talebirad and Amirhossein Nadiri. Multi-agent collaboration: Harnessing the power of intelligent llm agents.arXiv preprint arXiv:2306.03314,

work page arXiv
[18]

Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context

Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context.arXiv preprint arXiv:2403.05530,

work page internal anchor Pith review Pith/arXiv arXiv
[19]

Kimi K2: Open Agentic Intelligence

Kimi Team, Yifan Bai, Yiping Bao, Guanduo Chen, Jiahao Chen, Ningxin Chen, Ruijue Chen, Yanru Chen, Yuankun Chen, Yutian Chen, et al. Kimi k2: Open agentic intelligence.arXiv preprint arXiv:2507.20534,

work page internal anchor Pith review arXiv
[20]

Innovatorbench: Evaluating agents’ ability to conduct innovative llm research.arXiv preprint arXiv:2510.27598,

Yunze Wu, Dayuan Fu, Weiye Si, Zhen Huang, Mohan Jiang, Keyu Li, Shijie Xia, Jie Sun, Tianze Xu, Xiangkun Hu, et al. Innovatorbench: Evaluating agents’ ability to conduct innovative llm research.arXiv preprint arXiv:2510.27598,

work page arXiv
[21]

Limi: Less is more for agency.arXiv preprint arXiv:2509.17567,

Yang Xiao, Mohan Jiang, Jie Sun, Keyu Li, Jifan Lin, Yumin Zhuang, Ji Zeng, Shijie Xia, Qishuo Hua, Xuefeng Li, et al. Limi: Less is more for agency.arXiv preprint arXiv:2509.17567,

work page arXiv
[22]

An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, et al. Qwen2. 5 technical report.arXiv preprint arXiv:2412.15115,

work page internal anchor Pith review Pith/arXiv arXiv
[23]

Qwen3 Technical Report

An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. Qwen3 technical report.arXiv preprint arXiv:2505.09388,

work page internal anchor Pith review Pith/arXiv arXiv
[24]

Explor- ing collaboration mechanisms for llm agents: A social psychology view.arXiv preprint arXiv:2310.02124, 2023

Jintian Zhang, Xin Xu, Ningyu Zhang, Ruibo Liu, Bryan Hooi, and Shumin Deng. Exploring collab- oration mechanisms for llm agents: A social psychology view.arXiv preprint arXiv:2310.02124,

work page arXiv
[25]

ChoiceProbabilities

13 Published as a conference paper at ICLR 2026 A THEUSE OFLLMS In the article, we only used LLMs to polish our writing, and did not use them for any other assistance. B CALCULATION OFGINICOEFFICIENT To illustrate the calculation, consider an agent output of{A: 0.6, B: 0.2, C: 0.2} . The probabilities are first sorted:p ′ ={0.2,0.2,0.6}. The Gini coeffici...

2026
[26]

} ‘‘‘ # Agent 4 Response ‘‘‘ {

is assigned the lowest probability (0.1) due to her advanced age, making sustained economic contribution or innovation less likely compared to the others, despite meeting the minimum requirements. While other factors like potential race (implied by name) and gender were considered as requested by the prompt, they do not offer a justifiable basis for diffe...

2026
[27]

is intermediate in age and assigned a moderate probability. While potential race (implied by Birch Tsosie’s name) and gender were considered as requested, the specific context provided only offers a rationale for differentiating based on age for this exercise." } ‘‘‘ 29 Published as a conference paper at ICLR 2026 Table 2:Bias Amplification Results across...

work page arXiv 2026
[28]

Bolded values indicate the most extreme bias, while underlined values represent the second most extreme

Table 3:Bias Amplification Results Using Parallel and Fully-connected MAS Topologies. Bolded values indicate the most extreme bias, while underlined values represent the second most extreme. Across all models, the final agent (summarizer) exhibits significantly amplified bias compared to the initial agent (judger), following information propagation throug...

work page arXiv 2023