Research map / Large language models / Instruction and fine-tuning
Bi-directional Bias Attribution: Debiasing Large Language Models without Modifying Prompts
Yujie Lin, Kunquan Li, YiXuan Liao, Xiaoxin Chen, Jinsong Su
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add FBA/BBA neuron intervention | Llama-3.1-8B · StereoSet Gender | ICAT: 45.31 -> 62.50 | Improved |
| add FBA intervention | Llama-3.1-8B · StereoSet Religion | ICAT: 78.48 -> 91.14 | Improved |
| add FBA/BBA intervention vs prompt methods | Llama-3.1-8B · BBQ ambiguous vs disambiguated | Acc_dis: 73.02 (Self-Debiasing) vs 88.87 (FBA) | Improved |
| add BBA intervention | Llama-3.1-8B · WinoBias | Gap: 25.26 -> 1.02 | Improved |
| add FBA intervention | Llama-3.2-3B · WinoBias | Gap: 91.42 -> 18.68 | Improved |
| add BBA intervention | Llama-3.1-8B · Bias-in-Bios | Gap-TPR: 6.05 -> 1.39 | Improved |
| add FBA intervention | Llama-3.1-8B · Bias-NLI | P_neutral: 81.8 -> 99.1 | Improved |
| apply IG2 neuron suppression instead | Llama-3.2-3B · StereoSet Gender | LMS: 97.66 -> 66.41 | Hurt |
| replace stereotype cue selection with first-word-of-group cues (ablation) | FBA, Llama-3.1 · StereoSet | SS | Hurt |
| replace attribution with random neuron selection (ablation) | FBA, Llama-3.1 · StereoSet | SS/LMS | Hurt |
| intervene on first-layer instead of last-layer neurons (ablation) | FBA, Llama-3.1 · StereoSet Gender | ICAT: 62.50 -> 48.44 | Hurt |
| intervene on first-layer neurons (nationality) (ablation) | FBA, Llama-3.1 · StereoSet Nationality | ICAT: 68.19 -> 68.19 | No clear effect |
Closest papers by meaning
- The Devil is in the Neurons: Interpreting and Mitigating Social Biases in Language Models (ICLR 2024 · similarity 0.91)
- Debiasing Algorithm through Model Adaptation (ICLR 2024 · similarity 0.90)
- Biases Mitigation and Expressiveness Preservation in Language Models: A Comprehensive Pipeline (Student Abstract) (AAAI 2024 · similarity 0.88)
- UniBias: Unveiling and Mitigating LLM Bias through Internal Attention and FFN Manipulation (NeurIPS 2024 · similarity 0.87)
- Prompting Fairness: Integrating Causality to Debias Large Language Models (ICLR 2025 · similarity 0.87)
- Evaluating the Efficacy of Prompting Techniques for Debiasing Language Model Outputs (Student Abstract) (AAAI 2024 · similarity 0.87)
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.
