Research map / Large language models / Language, safety and interpretability
Silenced Biases: The Dark Side LLMs Learned to Refuse
Rom Himelstein, Amit LeVi, Brit Youngmann, Yaniv Nemcovsky, Avi Mendelson
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 refusal steering | Llama-2-7B · SBB | ASR (%): 19.63 -> 100 | Improved |
| add refusal steering | Qwen-14B · SBB | ASR (%): 12.85 -> 94.33 | Improved |
| add refusal steering | Llama-3.1-8B · SBB | ASR (%): 92.25 -> 99.98 | Improved |
| add refusal steering | Gemma-7B · SBB | ASR (%): 48.73 -> 98.92 | Improved |
| add refusal steering | Llama-2-7B · SBB | runtime (GPU hours): 15.82 -> 23.06 | Hurt |
| add refusal steering | Llama-3.1-8B-Instruct · BBQ ambiguous questions | refusal rate (%): 73 -> 69 | Improved |
| add refusal steering, with refusal option removed from BBQ | Llama-3.1-8B-Instruct · BBQ without refusal option | refusal rate (%): 69 -> 37 | Improved |
| add refusal steering | Llama-3.1-8B-Instruct · BBQ Race bias score | bias score: -1.15 -> -0.48 | No clear effect |
| add refusal steering | Llama-3.1-8B-Instruct · BBQ Appearance bias score | bias score: 15.51 -> 15.13 | No clear effect |
| vary refusal direction variant (10 sampled directions) (ablation) | refusal steering · SBB | Jaccard similarity of responses: minimum 0.75 | No clear effect |
| add adversarial-suffix jailbreak instead of refusal steering | Llama-2-7B, Gemma-7B, Qwen-7B · SBB | DPD consistency across runs | Hurt |
Closest papers by meaning
- Is Your Model Fairly Certain? Uncertainty-Aware Fairness Evaluation for LLMs (ICML 2025 · similarity 0.91)
- FairMT-Bench: Benchmarking Fairness for Multi-turn Dialogue in Conversational LLMs (ICLR 2025 · similarity 0.90)
- B-score: Detecting biases in large language models using response history (ICML 2025 · similarity 0.89)
- All Should Be Equal in the Eyes of LMs: Counterfactually Aware Fair Text Generation (AAAI 2024 · similarity 0.89)
- Prompting Fairness: Integrating Causality to Debias Large Language Models (ICLR 2025 · similarity 0.89)
- LIDAO: Towards Limited Interventions for Debiasing (Large) Language Models (ICML 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.
