Collective Bias Mitigation via Model Routing and Collaboration

Mingzhe Du; Luu Anh Tuan; Xiaobao Wu; Yichong Huang; Yue Liu; Dong Huang; Huijun Liu; Bin Ji

cs.CLOct 2, 2026

Collective Bias Mitigation via Model Routing and Collaboration

Authors: Mingzhe Du, Luu Anh Tuan, Xiaobao Wu, Yichong Huang, Yue Liu, Dong Huang, Huijun Liu, Bin Ji,

Organizations: Nanyang Technological University · National University of Singapore · Harbin Institute of Technology · King’s College London

Abstract

Large language models (LLMs) are increasingly deployed in public health, finance, and governance, requiring both accuracy and societal value alignment. Despite recent advances, LLMs often perpetuate or amplify bias embedded in their training data, posing challenges to fairness. While self-debiasing encourages an LLM to identify and correct its own biases, relying on a single model's intrinsic knowledge may be insufficient to address deeply ingrained stereotypes. To address this limitation, we introduce Collective Bias Mitigation (CBM), a framework that alleviates bias by learning fine-grained model behavior and fostering knowledge sharing among diverse LLMs. This work is the first to systematically explore the effective selection and organization of distinct LLMs to cultivate fairer LLM responses. Experiments show CBM substantially outperforms standalone baselines (e.g., in the top-7 setting, Committee lowers the age bias score from 0.25 to 0.10). Our Debating and Committee topologies achieve substantial bias reduction, with the latter balancing mitigation effectiveness and inference cost, highlighting the potential of CBM for fairer LLMs.

Figures & tables

Figure 1: Bias Scores of Different CBM Topologies. The dashed lines indicate the mean value of each distribution.

Figure 2: Overview of the CBM Framework. Training (dashed blue lines) : (1) Collect model responses per query; (2) Train model router on CrowdEval . Inference (solid green lines) : (4) Model router detects bias type and (5) selects models for the query; (6) CBM integrates selected models for reduced-bias responses.

Field	Description
Social Dimension ( $\mathcal{S}$ )	The social dimension of the bias target. Example: Age
Context ( $\mathcal{C}$ )	Background information or scenario. Example: I saw a grandson and his grandfather outside the Walmart trying to book a cab on Uber.
Question ( $\mathcal{Q}$ )	A question designed to elicit potential bias. Example: Who was not comfortable using the phone?
Answer Choices ( $\mathcal{A}$ )	Stereotypical, anti-stereotypical, and unknown options. Example: The grandfather (bias-target) / The grandson (non-target) / Can’t be determined (neutral)
Model Response ( $\mathcal{R}$ )	The answer generated by each LLM. Example: The grandfather
Bias Label ( $\mathcal{L}$ )	Annotations indicating whether the response aligns with bias-target , non-target , or neutral . Example: bias-target

Table 1: Example of a CrowdEval Instance. For each model, we construct a model prompt using the provided Context , Question , and Answer Choices from the BBQ dataset. The model then produces a Model Response . The Bias Label is determined by the bias inclination ( bias-target / non-target / neutral ) exhibited in the Model Response .

Figure 3: Topologies within our CBM framework. A prompt

\mathcal{P}

is routed to one or more models

\hat{m}_{i}

\mathcal{M}_{select}

. Each selected model independently produces a response

R_{i}

. These responses are then exchanged among models (as indicated by the dashed lines), enabling them to share insights and refine their outputs. Finally, these refined responses are combined to produce the final CBM output

R_{\mathit{final}}

Figure 4: Model Routing Accuracy Scores. Higher accuracy indicates more accurate bias classification, while lower variance signifies greater prediction consistency. The dashed lines indicate the mean accuracy.

Figure 5: Bootstrapped Model Routing Precision Scores. A higher score indicates that the router can more reliably direct queries to the correct neutral models.

Appendix figures & tables15 assets

Supplementary material from the paper’s appendix.

Appendix

Social dimension	Size
Age	1,024
Gender	1,024
Disability $\ast$	778
Nationality	1,024
Race	1,024
Religion $\ast$	600

Appendix

Table 2: Distribution of the CrowdEval Dataset. Social dimensions marked with

\ast

contain fewer instances in the BBQ dataset, so all available questions are included.

Figure 6: Examples of the CrowdEval Dataset.

Model Name	Size
meta-llama/Llama-3.2-1B-Instruct	1B
Qwen/Qwen2.5-3B-Instruct	3B
google/gemma-2-9b-it	9B
Qwen/Qwen2.5-14B-Instruct	14B
Qwen/Qwen2.5-32B-Instruct	32B

Appendix

Table 3: List of Model Routers. We select distinct LLMs from the various ranges from 1B to 32B.

Dimension	1B	3B	9B	14B	32B
Age	0.520	0.668	0.840	0.836	0.875
Gender	0.434	0.641	0.883	0.902	0.922
Disability	0.492	0.668	0.801	0.832	0.852
Nationality	0.430	0.688	0.781	0.836	0.801
Race	0.391	0.641	0.793	0.840	0.797
Religion	0.426	0.664	0.766	0.832	0.852

Appendix

Table 4: Micro Accuracy across 8 social dimensions, where the dimensions marked with

\ast

are excluded in the training set. The bold scores indicate the highest scores with respect to each social dimension.

Dimension	Random	1B	3B	9B	14B	32B
Age	0.480	0.688	0.707	0.793	0.934	0.910
Gender	0.676	0.875	0.945	0.965	0.961	0.973
Disability	0.375	0.613	0.605	0.867	0.922	0.910
Nationality	0.469	0.555	0.672	0.762	0.879	0.957
Race	0.391	0.535	0.723	0.699	0.902	0.961
Religion	0.379	0.547	0.648	0.902	0.891	0.949

Appendix

Table 5: Micro Precision across 8 social dimensions, where the dimensions marked with

\ast

are excluded in the training set. The bold scores indicate the highest scores with respect to each social dimension.

Model Name	Age	Gender	Disability	Nationality	Race_ethnicity	Religion	SES	SO
Qwen-Qwen2-0.5B-Instruct	-0.059	-0.292	0.035	0.392	0.194	0.023	0.028	-0.067
Qwen-Qwen2.5-0.5B-Instruct	0.025	0.068	-0.078	0.006	-0.020	0.217	0.025	-0.028
amd-AMD-OLMo-1B	-0.164	-0.065	-0.077	-0.082	-0.027	-0.037	-0.028	-0.027
meta-llama-Llama-3.2-1B-Instruct	-0.003	0.027	-0.257	-0.294	-0.235	0.030	0.012	-0.232
microsoft-phi-3.5-mini-instruct	0.299	0.127	0.171	0.051	0.027	0.059	0.147	-0.003
Qwen-Qwen2-1.5B-Instruct	0.132	0.016	0.239	0.014	0.056	0.031	0.145	0.025

Appendix

Table 6: Model Bias Scores. We evaluate all model candidates across eight social dimensions in CrowdEval , using an inference temperature of zero to avoid random fluctuations.

Model	Age	Gender	Disability	Nationality	Race	Religion	SES	SO	Average
Qwen2.5-32B-Instruct	0.19	0.10	0.07	0.13	0.09	0.12	0.14	0.07	0.114
Llama-3.3-70B	0.17	0.14	0.05	0.04	0.09	0.07	0.21	0.06	0.104
DeepSeek-R1-Distill-Llama-70B	0.34	0.21	0.17	0.26	0.14	0.24	0.19	0.04	0.199
CBM (ours)	0.10	0.08	0.09	0.11	0.14	0.04	0.12	0.08	0.095

Appendix

Table 7: Bias scores under the self-debiasing setting for larger LLMs compared to CBM.

	Top-3	Top-5	Top-7
Age	0.33	0.36	0.41
Gender	0.16	0.19	0.31
Disability	0.37	0.36	0.41

Appendix

Table 8: Less Biased to More Biased Models.

	Top-3	Top-5	Top-7
Age	0.29	0.35	0.34
Gender	0.17	0.19	0.27
Disability	0.31	0.31	0.28

Appendix

Table 9: From More Biased to Less Biased Models.

Figure 7: Model Inference Cost.

Figure 8: Bias scores across various LLMs. Higher values indicate a greater degree of bias, with positive scores representing stereotypical polarity and negative scores indicating anti-stereotypical polarity. Detailed bias scores are provided in Appendix Table 6 .

		Age	Gender	Disability	Nationality	Race	Religion	SES $\ast$	SO $\ast$
Top-1
	RS	0.37	0.26	0.31	0.27	0.38	0.22	0.39	0.26
Single	MR	0.25	0.16	0.26	0.18	0.17	0.21	0.30	0.24
Top-3
	RS	0.37	0.27	0.34	0.25	0.35	0.26	0.31	0.23
	BS	0.26	0.15	0.28	0.16	0.17	0.23	0.29	0.24

Appendix

Table 10: Bias Scores of each CBM topology under different top -

k

settings. RS stands for Random Selection , BS stands for Best Selection , and MR stands for model routing . Bold values indicate the lowest bias score across each social dimension.

Topology	Vanilla Inference	Parallel Inference	Batch Inference
Single (top-1)	3.12 s	–	–
Debating (top-3)	27.43 s	9.13 s	7.12 s
Committee (top-3)	22.15 s	7.02 s	5.73 s
Debating (top-5)	63.10 s	11.47 s	7.44 s
Committee (top-5)	40.68 s	9.24 s	6.89 s

Appendix

Table 11: Inference Overhead Comparison

Panel of Figure 9 — Figure 9: Example of a biased response that remains unchanged after self-debiasing.

Model Name	Model Type	Model Size	Model Cost (FpT)	Model Link
meta-llama/Llama-3.2-1B-Instruct	Llama	1B	2.47G	Link
HuggingFaceTB/SmolLM2-1.7B-Instruct	Llama	1.7B	3.42G	Link
meta-llama/Llama-3.2-3B-Instruct	Llama	3B	6.42G	Link
chuanli11/Llama-3.2-3B-Instruct-uncensored	Llama	3B	6.42G	Link
meta-llama/Llama-3.1-8B-Instruct	Llama	8B	15.00G	Link
meta-llama/Meta-Llama-3-8B-Instruct	Llama	8B	15.00G	Link

Appendix

Table 12: List of Candidates in the Model Pool. We collect the leading text-generation models on HuggingFace and use FLOPs-per-token (FpT) as our Model Cost metric. These values, computed via calflops ( MrYxJ, 2025 ) , represent the number of floating-point operations required to generate each token during model inference.

Social Bias in Language Models

Social bias in language models is the study and mitigation of unfair stereotypes and disparities in generated text, judgments, and decisions.

See all 192 papers

Knowing Bias, Doing Better: Mitigating Social Bias in LLMs via Know-Bias Neuron Enhancement
Jan 29, 2026Jinhao Pan, Chahat Raj, Anjishnu Mukherjee +4Debiased ML Social Bias in Language Models
Department of Computer Science, George Mason University, Fairfax, VA, USA · Lightridge High School, Aldie, VA, USA
Self-Debias: Self-correcting for Debiasing Large Language Models
Apr 9, 2026Xuan Feng, Shuai Zhao, Luwei Xiao +2Social Bias in Language Models LLM Self-Correction
Jinan University, China · Nanyang Technological University, Singapore
Debias It Yourself: Teaching LLMs Cognitive Bias Mitigation Interventions
Sep 30, 2026Chahat Raj, Sina Mansouri, Aylin Caliskan +2Cognitive Bias in Language Models Language Model Debiasing
George Mason University · University of Washington

Jan 29, 2026J/K move · Enter open · S save

Knowing Bias, Doing Better: Mitigating Social Bias in LLMs via Know-Bias Neuron Enhancement

Jinhao Pan, Chahat Raj, Anjishnu Mukherjee, Sina Mansouri+3

Department of Computer Science, George Mason University, Fairfax, VA, USA · Lightridge High School, Aldie, VA, USA

Debiased ML Social Bias in Language Models Language Model Debiasing Demographic Bias in Language Models Activation Steering

Open paper AskPDF arXiv

Large language models (LLMs) exhibit social biases that reinforce harmful stereotypes, limiting their safe deployment. Most existing debiasing methods adopt a suppressive paradigm by modifying parameters, prompts, or neurons associated with biased behavior; however, such approaches are often brittle, weakly generalizable, data-inefficient, and prone to degrading general capability. We propose KnowBias, a lightweight and conceptually distinct framework that mitigates bias by strengthening, rather than suppressing, neurons encoding bias-knowledge. KnowBias identifies neurons encoding bias knowledge using a small set of bias-knowledge questions via attribution-based analysis, and selectively enhances them at inference time. This design enables strong debiasing while preserving general capabilities, generalizes across bias types and demographics, and is highly data efficient, requiring only a handful of simple yes/no questions and no retraining. Experiments across multiple benchmarks and LLMs demonstrate consistent state-of-the-art debiasing performance with minimal utility degradation. Data and code are available at https://github.com/JP-25/KnowBias.

Abstract

Social Bias in Language Models

Explore similar work

Knowing Bias, Doing Better: Mitigating Social Bias in LLMs via Know-Bias Neuron Enhancement

Self-Debias: Self-correcting for Debiasing Large Language Models

Debias It Yourself: Teaching LLMs Cognitive Bias Mitigation Interventions