Research map / Architectures and efficiency / RNNs and normalization
Debiasing Algorithm through Model Adaptation
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.
Causal tracing locates mid-upper MLP layers carrying gender bias in LLaMA models, and DAMA projects their output weights onto the orthogonal complement of a stereotype subspace to reduce bias while preserving perplexity and downstream accuracy.
Problem
Large language models encode gender stereotypes from training data (e.g., associating professions with a gender), causing biased pronoun predictions and uneven coreference performance; the goal is to detect which components carry this bias and edit weights to reduce it without hurting general perfor
Approach
Use causal tracing with a linear fit of pronoun probability differences against annotated stereotypical and factual gender scores to locate bias-mediating MLP layers, then edit them by multiplying the FF output matrix by a PLS-derived orthogonal projection that removes the stereotype subspace.
Key results
- On LLaMA 7B, DAMA reduces the stereotypical coefficient a_s from 0.235 to -0.005 and WinoBias delta S from 40.3% to 31.5%, while StereoSet ICAT rises from 53.7 to 58.5.
- DAMA on LLaMA 7B raises perplexity from 26.1 to 28.9 and changes ARC-E significantly (68.3 vs 69.1), with other downstream tasks largely preserved.
- DAMA reduces bias across 13B, 30B and 65B models, e.g., 65B WinoBias delta S drops from 35.7% to 27.2%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Project selected MLP weights onto the subspace orthogonal to stereotypical directions | LLaMA 7B · Gendered prompts | a_s: 0.235 -> -0.005 | Helped |
| Project selected MLP weights onto the subspace orthogonal to stereotypical directions | LLaMA 7B · WinoBias | ΔS: 40.3% -> 31.5% | Helped |
| Project selected MLP weights onto the subspace orthogonal to stereotypical directions | LLaMA 7B · StereoSet gender intrasentence | ICAT: 53.7 -> 58.5 | Helped |
| Project selected MLP weights onto the subspace orthogonal to stereotypical directions | LLaMA 7B · Wikipedia 103 | PPL: 26.1 -> 28.9 | Hurt |
| Project selected MLP weights onto the subspace orthogonal to stereotypical directions | LLaMA 7B · ARC-C | accuracy: 42.2 -> 41.8 | No clear effect |
| Project selected MLP weights onto the subspace orthogonal to stereotypical directions | LLaMA 7B · ARC-E | accuracy: 69.1 -> 68.3 | Hurt |
| Project selected MLP weights onto the subspace orthogonal to stereotypical directions | LLaMA 7B · MMLU | accuracy: 30.3 -> 30.8 | No clear effect |
| Project selected MLP weights onto the subspace orthogonal to stereotypical directions | LLaMA 13B, LLaMA 30B, and LLaMA 65B · Gendered prompts | a_s: 0.270 -> 0.148; 0.265 -> 0.105; 0.249 -> 0.185 | Helped |
| Project selected MLP weights onto the subspace orthogonal to stereotypical directions | LLaMA 13B, LLaMA 30B, and LLaMA 65B · WinoBias | ΔS: 35.7% -> 31.1%; 36.0% -> 26.7%; 35.7% -> 27.2% | Helped |
| Expand the number of edited layers and projection dimensions (ablation) | DAMA on LLaMA 7B · Gendered prompts | a_s | Helped |
| Expand the number of edited layers and projection dimensions (ablation) | DAMA on LLaMA 7B · Wikipedia 103 | PPL | Hurt |
| Apply DAMA to the two last model layers (ablation) | DAMA · Gender bias evaluation | a_s | Hurt |
Surprising findings
- Last-layer MLPs appear to counter the gender bias in the original model, and intervening on them worsens both bias and perplexity.
- MEMIT has minor effect on bias measures and does not generalize to unseen professions or other bias types.
Limitations
- Focused only on gender bias in English and on LLaMA-family models
- Coreference accuracy drops (e.g., LLaMA 65B 73.3% to 71.1%) due to weakened stereotypical cues
- Perplexity increases for 7B model after intervention
Experimental setup
Datasets: Bolukbasi et al. professions list (factual and stereotypical scores) WinoBias StereoSet Wikitext-103 OpenBookQA ARC MMLU
Benchmarks: WinoBias StereoSet Wikitext-103 perplexity OpenBookQA ARC-Challenge ARC-Easy MMLU
Metrics: linear coefficients a_s, a_f, b, R^2 WinoBias accuracy, delta S, delta G StereoSet lms, ss, ICAT perplexity accuracy
Baselines: MEMIT LoRA FT LLaMA 7B
Builds on: MEMIT causal tracing LoRA
Closest papers by meaning
- Biases Mitigation and Expressiveness Preservation in Language Models: A Comprehensive Pipeline (Student Abstract) (AAAI 2024 · similarity 0.91)
- Auto-Search and Refinement: An Automated Framework for Gender Bias Mitigation in Large Language Models (NeurIPS 2025 · similarity 0.90)
- Are Models Biased on Text without Gender-related Language? (ICLR 2024 · similarity 0.89)
- Mitigating Social Bias in Large Language Models: A Multi-Objective Approach Within a Multi-Agent Framework (AAAI 2025 · similarity 0.88)
- Prompting Fairness: Integrating Causality to Debias Large Language Models (ICLR 2025 · similarity 0.88)
- Evaluating the Efficacy of Prompting Techniques for Debiasing Language Model Outputs (Student Abstract) (AAAI 2024 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.