Acmite: Mitigating Gender Bias in LLMs through Concept-Guided Mutual Information

Tian Lan; Xiaoqing Cheng; Han Zhang; Jiang Li

cs.CLOct 1, 2026

Acmite: Mitigating Gender Bias in LLMs through Concept-Guided Mutual Information

Authors: Tian Lan, Xiaoqing Cheng, Han Zhang, Jiang Li

Organizations: Kyoto University · Zhejiang University · Shanghai Jiao Tong University · Inner Mongolia University

Abstract

Large language models (LLMs) can reproduce social stereotypes from their training data, motivating extensive research on model debiasing. However, existing methods often rely on explicit biased examples or predefined group-term substitutions, making them sensitive to wording and less effective at capturing stereotype concepts shared across diverse contexts. More importantly, they typically suppress biased outputs without explicitly modeling the statistical dependence between model outputs and the underlying stereotype concepts. We propose Acmite, a lightweight concept-guided framework for targeted and selective debiasing. Acmite represents stereotypes as structured semantic concepts and uses maximal marginal relevance (MMR) to select diverse concepts for debiasing. Inspired by mutual information minimization, it approximates this dependence with token-level KL divergence while preserving task semantics. A lightweight LoRA adapter is trained with the base model frozen and activated at inference time only when the input is sufficiently similar to stereotype-related concepts; otherwise, the original model is used directly. We evaluate Acmite on BBQ, CrowS-Pairs, and StereoSet, and assess general capability preservation on ARC-Challenge, GSM8K, and PIQA. Experiments across three LLMs show that Acmite effectively mitigates gender bias across complementary evaluation formats while maintaining competitive performance on bias-unrelated tasks. Anonymous code and data are available at https://anonymous.4open.science/r/Acmite-18E2/.

Figures & tables

Phi-4-mini-Instruct
Method	BBQ				CrowS-Pairs	StereoSet
Method	Amb. Acc $\uparrow$	Amb. $S$ $\rightarrow$ 0	Dis. Acc $\uparrow$	Dis. $S$ $\rightarrow$ 0	SS $\rightarrow$ 50	LMS $\uparrow$	SS $\rightarrow$ 50	ICAT $\uparrow$
Base	92.37	-1.59	89.42	1.18	59.06	93.92	67.45	61.14
+BiasDPO	95.66	-1.09	86.85	1.45	59.38	93.46	65.97	63.61
+SynDebias	87.09	-2.54	89.25	1.34	59.38	94.24	68.32	59.70
+GenderAlign	87.17	-0.49	62.13	0.77	60.62	91.62	67.02	60.43

Table 1: Overall gender debiasing results. Bold and underlined values indicate the best and second-best results among debiasing methods, respectively; Base is excluded from ranking.

Method	BBQ				CrowS-Pairs	StereoSet
Method	Amb. Acc $\uparrow$	Amb. $S$ $\rightarrow$ 0	Dis. Acc $\uparrow$	Dis. $S$ $\rightarrow$ 0	SS $\rightarrow$ 50	LMS $\uparrow$	SS $\rightarrow$ 50	ICAT $\uparrow$
Phi-4-mini-Instruct
Base	92.37	-1.59	89.42	1.18	59.06	93.92	67.45	61.14
Acmite (+22.9%)	89.54 (-3.1%)	-0.46 (+71.1%)	90.09 (+0.7%)	0.58 (+50.8%)	55.63 (+37.9%)	93.46 (-0.5%)	64.39 (+17.5%)	66.54 (+8.8%)
$w/o$ $\mathcal{L}_{\mathrm{MI}}$ (+2.4%)	93.34 (+1.1%)	-1.52 (+4.4%)	89.32 (-0.1%)	1.61 (-36.4%)	55.00 (+44.8%)	94.90 (+1.0%)	67.05 (+2.3%)	62.54 (+2.3%)
$w/o$ $\mathcal{L}_{\mathrm{dir}}$ (+18.3%)	92.38 (+0.01%)	-1.06 (+33.3%)	89.07 (-0.4%)	0.55 (+53.4%)	55.63 (+37.9%)	93.72 (-0.2%)	64.92 (+14.5%)	65.75 (+7.5%)

Table 2: Ablation results on Phi-4-mini-Instruct. Green and red percentages denote improvement and degradation over Base; the percentage beside each method is the mean across all eight metrics. Rankings exclude Base.

Panel of Table 3 — Table 3: A representative open-ended case. The base model introduces an unsupported division of responsibilities, whereas Acmite preserves the equal contribution specified in the prompt.

Phi-4-mini-Instruct
Method	ARC $\uparrow$	GSM8K $\uparrow$	PIQA $\uparrow$	Avg. $\uparrow$
Base	61.26	80.89	77.15	73.10
+BiasDPO	61.69	81.88	76.55	73.37
+SynDebias	61.01	80.74	77.10	72.95
+GenderAlign	56.06	80.89	73.83	70.26
+BiasEdit	37.71	76.50	59.09	57.77

Table 4: General capability results. Avg. is the arithmetic mean of ARC, GSM8K, and PIQA. Bold and underlined values indicate the best and second-best results among debiasing methods, respectively; Base is excluded from ranking.

Panel of Figure 1 — Figure 1: Fairness–utility trade-offs relative to Base. Both axes report percentage changes; mean bias improvement follows the direction-aware calculation in Table 2 . Base is excluded, and dashed lines show the Pareto frontiers.

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

Hyperparameter	Value
Optimizer	AdamW
Learning rate	$1\times 10^{-4}$
Epochs	4
Effective batch size	16
LoRA rank / scaling	16 / 32
$\lambda_{\mathrm{MI}}$ schedule	$[1,1,1,2]$ (Qwen3);

Appendix

Table 5: Hyperparameters used for Acmite training and routing.

Panel of Table 6 — Table 6: Prompt for stereotype concept extraction from FSB and GEST.

Panel of Table 7 — Table 7: Prompt for auditing model-specific neutral responses.

Panel of Table 8 — Table 8: Prompt for generating fair-reference statements.

Panel of Table 9 — Table 9: An occupational-role case. Red marks unsupported gender attribution; green marks the corresponding gender-neutral continuation.

Panel of Table 10 — Table 10: A promotion-and-childcare case. Acmite avoids assigning mobility according to gender.

Panel of Table 11 — Table 11: A tied-hiring case. Acmite introduces job-relevant criteria without attaching them to gender.

Explore similar work

Knowing Bias, Doing Better: Mitigating Social Bias in LLMs via Know-Bias Neuron Enhancement
Jan 29, 2026Jinhao Pan, Chahat Raj, Anjishnu Mukherjee +4Large Language Model Bias Debiasing
Department of Computer Science, George Mason University, Fairfax, VA, USA · Lightridge High School, Aldie, VA, USA
BiasGym: A Simple and Generalizable Framework for Analyzing and Removing Biases through Injection
Aug 12, 2025Sekh Mainul Islam, Nadav Borenstein, Siddhesh Milind Pawar +3Large Language Model Bias Biases
University of Copenhagen
DebiasRAG: A Tuning-Free Path to Fair Generation in Large Language Models through Retrieval-Augmented Generation
May 15, 2026Rui Chu, Bingyin Zhao, Thanh Quoc Hung Le +6Large Language Model Bias Debiasing

Jan 29, 2026, cs.AIJ/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

Large Language Model Bias Debiasing Biases Suppression

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 \textbf{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

Explore similar work

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

BiasGym: A Simple and Generalizable Framework for Analyzing and Removing Biases through Injection

DebiasRAG: A Tuning-Free Path to Fair Generation in Large Language Models through Retrieval-Augmented Generation