PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Tian Lan, Xiaoqing Cheng, Han Zhang, Jiang Li

arXiv:2610.01696v1cs.CL

TL;DR

LLMs can reproduce social stereotypes, and prior debiasing often targets specific examples rather than modeling shared concepts and their dependence on outputs. Acmite uses concept-guided mutual-information optimization with selective activation; across three model families, experiments report gender-bias mitigation and competitive performance on bias-unrelated tasks.

  • Problem

    LLMs can inherit social biases from pretraining that influence predictions and generations, while prior work notes associations between demographic identities and stereotype concepts.

    Source: §1 · p. 1

  • Method

    Acmite retrieves shared stereotype concepts, approximates conditional mutual information with token-level KL divergence, and selectively activates debiasing for concept-similar inputs.

    Source: §1 · p. 2

  • Results

    Across three model families, Acmite mitigates gender bias on BBQ, CrowS-Pairs, and StereoSet while maintaining competitive performance on ARC-Challenge, GSM8K, and PIQA.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Retrieval-based routing limits intervention to concept-relevant inputs, reducing unnecessary changes to bias-unrelated inputs.

    Source: §1 · p. 2

    The evaluation covers gender bias only and represents gender with binary categories, leaving other attributes and non-binary and other gender identities unassessed.

    Source: Limitations · pp. 8–9

Abstract

from arXiv · show

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/.

1 Introduction

The introduction argues that varied expressions can share stereotype concepts, a gap that limits approaches targeting individual instances. Acmite uses concept-guided optimization and selective activation, and experiments report reduced gender bias with competitive general-task performance.

1 Introduction

  • Social biases can be inherited from pretraining and influence generations, potentially reinforcing unfairness and amplifying existing inequalities in high-impact domains.The introduction names hiring, education, and healthcare as examples.Source: §1 · p. 1
  • Existing methods target specific biased instances or constrain outputs with prompts, while varied expressions can share concepts that individual-expression optimization may miss.The introduction also notes sensitivity to training distributions and wording, and dependence of prompting methods on instruction following and generation stability.Source: §1 · pp. 1–2
  • Acmite retrieves shared stereotype concepts, optimizes conditional mutual information via token-level KL divergence with fair-reference alignment, and activates debiasing for sufficiently similar inputs.The framework is described as comprising stereotype concept retrieval, concept-guided optimization, and retrieval-guided selective activation.Source: §1 · p. 2
  • Across six benchmarks and three LLM families, Acmite mitigates gender bias while maintaining competitive performance on bias-unrelated tasks.The evaluation covers BBQ, CrowS-Pairs, StereoSet, ARC-Challenge, GSM8K, and PIQA.Source: §1 · p. 2

2 Related Work

The related work frames social bias as context-dependent associations between social attributes and model behavior, assessed through complementary benchmark formats. It organizes debiasing methods by intervention point and distinguishes Acmite through structured concepts shared across training and inference.

2.1 Social Bias in Large Language Models

  • Bias is framed as context-dependent associations between social attributes and model behavior, assessed through probability comparisons and contextual accuracy-and-bias measurements.The cited benchmarks include CrowS-Pairs, StereoSet, and BBQ, alongside resources for generation, multilingual settings, and concept sourcing.Source: §2.1 · p. 2

2.2 Debiasing Large Language Models

  • Prior methods intervene in data and training, model representations and parameters, or generation; Acmite instead uses structured stereotype concepts across training and inference.Acmite also guides optimization with fair references and selectively activates its adapter according to concept relevance.Source: §2.2 · p. 3

3 Method

Acmite has three stages: retrieve structured stereotype concepts, train an adapter to reduce stereotype-conditioned dependence while preserving task behavior, and route inputs to the adapter by similarity. Its retrieval stage diversifies concepts, and its optimization aligns stereotype-conditioned predictions with fair references.

3.1 Bias Concept Retrieval

  • MMR selects context-relevant, complementary stereotype concepts, discouraging near-duplicate descriptions while raw similarity remains the basis for relevance filtering and inference routing.The repository represents each concept by a social group, a self-contained stereotype description, and a semantic domain, omitting source-example wording.Source: §3.1 · p. 3

3.2 Concept-Guided Mutual Information Optimization

  • For training, model-specific neutral responses are checked for residual gender assumptions and revised only when needed, preserving task-relevant content as audited targets.The objective first preserves the audited response under the neutral branch, then approximates conditional mutual information with token-level KL divergence and aligns stereotype-conditioned predictions with fair references.Source: §3.2 · p. 4
  • The fair-reference alignment prevents the neutral branch from serving as the sole fairness target when it retains residual associations, and its stop-gradient keeps the reference fixed.The alignment term is zero when the stereotype-conditioned and fair-reference distributions agree, avoiding pressure toward increasingly strong reverse associations.Source: §3.2 · p. 4
  • Only low-rank adapter parameters are optimized; the base model remains unchanged, and the separate adapter can be activated selectively at inference.The training objective combines task-behavior preservation, mutual-information minimization, and fair-reference alignment.Source: §3.2 · p. 4

3.3 Retrieval-Guided Inference

  • At inference, raw cosine similarity determines whether the input enters a stereotype-related semantic region; the adapter route is used when activated, otherwise the frozen base model responds directly.The small adapter update avoids maintaining two full language models.Source: §3.3 · p. 4

4 Experiments

The experiments evaluate gender bias on three English benchmarks and capability preservation on three general-task benchmarks, comparing Acmite with the unmodified backbone and five debiasing baselines. They use three model families, a shared concept repository, and one inference threshold.

4.1 Evaluation Details

  • Gender bias is evaluated on BBQ, CrowS-Pairs, and StereoSet, while ARC-Challenge, GSM8K, and PIQA assess general capability preservation.BBQ tests stereotypical reasoning in ambiguous and disambiguated question-answering; CrowS-Pairs and StereoSet use model likelihoods, with higher accuracy, LMS, and ICAT preferred and BBQ S ideally zero.Source: §4.1 · pp. 4–5

4.2 Baselines

  • Acmite is compared with the unmodified backbone and five baselines: BiasDPO, SynDebias, GenderAlign, BiasEdit, and DSO.All methods use identical benchmark splits and metrics.Source: §4.2 · p. 5

4.3 Training Details

  • Main experiments use Phi-4-mini-Instruct, with cross-model evaluation on Qwen3-4B-Instruct-2507 and Llama-3.1-8B-Instruct; the repository has 3,953 concepts and 3,704 training contexts.A single inference threshold, τ = 0.60, is used across backbones and benchmarks without dataset-specific adjustment.Source: §4.3 · p. 5

5 Results and Analysis

Across the evaluated models, Acmite improves gender-bias measures across complementary benchmarks while retaining competitive general-task performance. Case studies and ablations illustrate selective generation behavior and the contributions of the optimization and routing choices.

5.1 Debiasing Results

  • On Phi-4-mini, Acmite improves BBQ, CrowS-Pairs, and StereoSet results; gains across Qwen3 and Llama-3.1-8B likewise span complementary bias measures.Phi-4-mini reaches closest-to-zero BBQ bias scores (−0.46 and 0.58), highest disambiguated accuracy (90.09), and best StereoSet SS and ICAT (64.39 and 66.54); CrowS-Pairs moves from 59.06 to 55.63.Source: §5.1 · pp. 5–6

5.2 General Capability Results

  • Across three model families, Acmite maintains competitive general-task performance while mitigating social bias.On Phi-4-mini, the reported average is 73.24 versus 73.10 for the base model; Qwen3 and Llama-3.1-8B also retain competitive capability.Source: §5.2 · p. 6

5.3 Case Study

  • In the representative startup case, Acmite assigns vision and execution jointly to both founders, avoiding the base model’s unsupported role division.Other selected cases show the same qualitative pattern: Acmite uses gender-independent descriptions and decision criteria where the base model introduces gendered identities or attributes absent from the prompt.Source: §5.3 · p. 7, Table 3 · p. 7

5.4 Ablation Studies

  • Full Acmite improves the eight-metric average by 22.9%, versus 18.3% without Ldir and 2.4% without LMI.The ablation reports that removing LMI sharply reduces the average gain, while the two terms have different metric trade-offs.Source: Table 2 · p. 7, §5.4 · p. 8
  • Retrieval-guided routing raises Phi-4-mini ARC-Challenge, GSM8K, and PIQA scores while activating the adapter on only 0.96% of ARC-Challenge and GSM8K inputs and 9.25% of PIQA inputs.Compared with always-on inference, the respective scores rise from 57.51, 77.71, and 69.21 to 61.60, 81.20, and 76.93.Source: §5.4 · p. 8
  • At τ = 0.55, activation rates are higher than at τ = 0.60, but the lower threshold does not produce stronger CrowS-Pairs or PIQA results.The comparison supports restricting adapter activation according to concept relevance rather than intervening on more inputs indiscriminately.Source: §5.4 · p. 8

6 Conclusion

Acmite organizes biased expressions into reusable stereotype concepts and uses concept-guided mutual information optimization with fair references and retrieval-based routing. Across six benchmarks and three model families, it reduces gender bias while maintaining competitive general capability.

  • Acmite treats biased expressions as reusable stereotype concepts and reduces their influence through concept-guided mutual information optimization, fair references, and retrieval-based routing.The conclusion characterizes this as moving beyond isolated correction targets.Source: §6 · p. 8
  • Low bias on question answering does not necessarily indicate weak stereotypical associations, motivating evaluation across complementary formats.Source: §6 · p. 8
  • Shared concepts extracted from diverse bias examples offer a viable concept-level alternative to instance-level preference-based debiasing.Source: §6 · p. 8

Limitations

The evaluation is limited to gender bias and represents gender with binary categories. Its effectiveness on other social attributes and its coverage of diverse gender identities remain unsettled or unrepresented.

Limitations

  • The experiments focus exclusively on gender bias, leaving Acmite's effectiveness on race, religion, age, and intersecting attributes to be established.The paper identifies these attributes as areas for future evaluation.Source: Limitations · p. 8
  • The evaluation represents gender with binary categories, so it does not capture bias involving non-binary and other gender identities.The paper says gender identities are diverse and cannot be fully represented by a binary division.Source: Limitations · p. 9
  • The authors call for extending the concept database and evaluation protocol to broader social attributes and more inclusive gender representations.Source: Limitations · pp. 8–9

Ethics Statement

The work aims to mitigate social bias, not endorse stereotypes. Its stereotype database may contain harmful statements, which are used only as research intervention signals; automated assistance may still introduce bias or errors.

Ethics Statement

  • The work aims to identify and mitigate social bias and does not endorse or reproduce stereotypical associations.Source: Ethics Statement · p. 9
  • The stereotype database is drawn from existing bias datasets and may contain offensive or harmful statements, used only as research intervention signals rather than factual descriptions.Source: Ethics Statement · p. 9
  • An external LLM assists with concept extraction, neutral-output auditing, and rewriting, but generated concepts and revisions may retain model-specific biases or errors.The paper reports using structured prompts and consistent processing criteria.Source: Ethics Statement · p. 9
  • The experiments involve neither human participants nor collection of new personal information.Source: Ethics Statement · p. 9

A.1 Hyperparameter Settings

The appendix reports a model-specific mutual-information weight schedule and describes prompts used for automated data construction, including concept extraction, response auditing, and fair-reference generation.

A.1 Hyperparameter Settings

  • Qwen3 uses the milder final-epoch MI-weight schedule [1, 1, 1, 2], while the other models use [1, 1, 1, 3].The authors attribute this choice to Qwen3's comparatively weak bias before debiasing.Source: §A.1 · p. 11

A.2 Prompts for Automated Data Construction

  • The appendix includes prompts and processing scripts for automated data construction with DeepSeek-V4-Flash.Source: §A.2 · p. 12
  • Separate extraction instructions address FSB's directly judged bias examples and GEST's gender-neutral sentences paired with fixed stereotype labels.The appendix also describes a five-demonstration neutral-response audit returning one ordered JSON result per input, and fair-reference generation with five demonstrations.Source: §A.2 · p. 12

B Additional Case Studies

The additional case studies show Acmite retaining task-relevant distinctions without assigning roles, mobility, childcare, or hiring criteria according to gender. The prompts use a qualitative protocol in which the needed continuation information does not depend on gender.

B Additional Case Studies

Loading 2610.01696v1…