Research map / Large language models / Language, safety and interpretability
Knowing Bias, Doing Better: Mitigating Social Bias in LLMs via Know-Bias Neuron Enhancement
Jinhao Pan, Chahat Raj, Anjishnu Mukherjee, Sina Mansouri, Bowen Wei, Shloka Yada, Ziwei Zhu
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 |
|---|---|---|---|
| enhance know-bias neurons at inference time instead of suppressing bias neurons or fine-tuning | Llama-3.2-3B-Instruct, Qwen-3-4B, Llama-3.1-8B · BBQ, CrowS-Pairs, StereoSet | total average rank: 1.8 / 2.1 / 2.3 (best among 8 methods) | Improved |
| use KnowBias instead of ReGiFT, PEFT, BiasEdit or SD for general capability | Llama-3.2-3B-Instruct · OBQA, COPA, ARC | accuracy drop: no drop of at least 5% for KnowBias vs many starred baselines | Improved |
| apply KnowBias to Llama-3.1-8B | Llama-3.1-8B-Instruct · COPA | accuracy: .7132 -> .6780 | Hurt |
| apply KnowBias to Llama-3.2-3B | Llama-3.2-3B-Instruct · ARC-E | accuracy: .6814 -> .7191 | Improved |
| use neurons identified from a single demographic dimension on other dimensions (ablation) | KnowBias on Llama-3.2-3B · SS-inter gender ICAT | ICAT: .7000 -> .8550 (race neurons) | Improved |
| use gender-derived neurons on religion bias (ablation) | KnowBias on Llama-3.2-3B · SS-inter religion ICAT | ICAT: .6196 -> .6883 | Improved |
| increase number of bias-knowledge questions beyond 45 (ablation) | KnowBias · StereoSet-Inter | ICAT | No clear effect |
| mix three question types instead of a single type (ablation) | KnowBias · BBQ, CS, SS | bias scores | Improved |
| use union of dimension-specific neuron sets instead of intersection or composite question (ablation) | KnowBias · BBQ, CS, SS | bias scores | Improved |
| use unified neuron set instead of dimension-specific neurons (ablation) | KnowBias · BBQ, CS, SS | bias scores | Improved |
| select know-bias neurons by attribution instead of random neurons of same count (ablation) | KnowBias · bias benchmarks | bias scores | Improved |
| use 5 bias concepts instead of 25 (ablation) | KnowBias · bias benchmarks | debiasing performance | No clear effect |
Closest papers by meaning
- The Devil is in the Neurons: Interpreting and Mitigating Social Biases in Language Models (ICLR 2024 · similarity 0.92)
- Prompting Fairness: Integrating Causality to Debias Large Language Models (ICLR 2025 · similarity 0.90)
- Biases Mitigation and Expressiveness Preservation in Language Models: A Comprehensive Pipeline (Student Abstract) (AAAI 2024 · similarity 0.89)
- Interpretable Debiasing of Vision-Language Models for Social Fairness (CVPR 2026 · similarity 0.89)
- Mitigating Social Bias in Large Language Models: A Multi-Objective Approach Within a Multi-Agent Framework (AAAI 2025 · similarity 0.89)
- Self-correcting for Debiasing Large Language Models (ICML 2026 · similarity 0.89)
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.
