Research map / Large language models / Language, safety and interpretability

Auto-Search and Refinement: An Automated Framework for Gender Bias Mitigation in Large Language Models

NeurIPS 2025 · Large language models › Language, safety and interpretability · approach: Bias measurement and mitigation · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes FaIRMaker, which auto-searches gradient-based debiasing trigger tokens (Fairwords), filters them, and trains a Llama3.2-3B seq2seq refiner to turn them into natural-language instructions that reduce gender bias while preserving utility across open and closed LLMs.

Problem

Gender bias mitigation in LLMs must be automated, applicable to closed-source models, and must not degrade general task performance; existing fine-tuning, editing, and manual preamble methods fail one or more of these.

Approach

Search debiasing trigger tokens with GCG on a preference dataset, filter them by measured effect, then train a seq2seq refiner that rewrites each trigger into a query-aware natural-language instruction (gender-related or neutral), so the module is model-independent and works on API models.

Key results

• On BBQ-gender, FaIRMaker reduces sDIS for Llama2-Alpaca from 1.066 to 0.224 and sAMB from 0.804 to 0.157.
• Response score on GA-test (DeepSeek judge) for Llama2-Alpaca rises from 3.71 to 4.07, and MMLU accuracy for Llama2-Alpaca rises from 33.49 to 37.93.
• Average processing time is about 1.5 seconds per query on a single A40 GPU.

Measured effects reported in the paper

Surprising findings

• Refined Fairwords without the refiner (vanilla Fairwords) almost fail to transfer to Qwen and GPT-3.5, while the refiner restores bias mitigation on black-box models.
• Auto-searched Fairwords are largely nonsensical token strings, yet they act as seeds whose interpreted emotional tones (urgency, frustration, anger) correlate with debiasing effect.

Limitations

• Focuses only on binary gender bias; other bias dimensions such as race or religion are not addressed.
• Refinement adds intermediate instructions that make some outputs more verbose and slightly lower scores on Dolly and Instruct Eval for stronger models.
• Relies on LLM-based judges (GPT-4, Llama3.1-8B) with only a 50-pair human check.

Experimental setup

Datasets: GenderAlign Alpaca BBQ Dolly MMLU HellaSwag Instruct Eval BPO Eval

Benchmarks: GenderAlign test (GA-test) BBQ-gender MMLU HellaSwag Dolly Eval Instruct Eval BPO Eval

Metrics: win-tie-loss rate sDIS sAMB response score accuracy time per query

Baselines: Intervention CF-D Desc-D MBIAS DPO

Builds on: Greedy Coordinate Gradient (GCG) Direct Preference Optimization Intervention prompting

Closest papers by meaning

• Debiasing Algorithm through Model Adaptation (ICLR 2024 · similarity 0.90)
• LIDAO: Towards Limited Interventions for Debiasing (Large) Language Models (ICML 2024 · similarity 0.89)
• Mitigating Social Bias in Large Language Models: A Multi-Objective Approach Within a Multi-Agent Framework (AAAI 2025 · similarity 0.89)
• All Should Be Equal in the Eyes of LMs: Counterfactually Aware Fair Text Generation (AAAI 2024 · similarity 0.89)
• Biases Mitigation and Expressiveness Preservation in Language Models: A Comprehensive Pipeline (Student Abstract) (AAAI 2024 · similarity 0.88)
• FairImagen: Post-Processing for Bias Mitigation in Text-to-Image Models (NeurIPS 2025 · similarity 0.87)