AI Paper Atlas

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

AAAI 2026 · Large language models › Language, safety and interpretability · approach: Uncertainty quantification for LLMs · 5 citations (OpenAlex)

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 changedWhereMeasuredEffect
add refusal steeringLlama-2-7B · SBBASR (%): 19.63 -> 100Improved
add refusal steeringQwen-14B · SBBASR (%): 12.85 -> 94.33Improved
add refusal steeringLlama-3.1-8B · SBBASR (%): 92.25 -> 99.98Improved
add refusal steeringGemma-7B · SBBASR (%): 48.73 -> 98.92Improved
add refusal steeringLlama-2-7B · SBBruntime (GPU hours): 15.82 -> 23.06Hurt
add refusal steeringLlama-3.1-8B-Instruct · BBQ ambiguous questionsrefusal rate (%): 73 -> 69Improved
add refusal steering, with refusal option removed from BBQLlama-3.1-8B-Instruct · BBQ without refusal optionrefusal rate (%): 69 -> 37Improved
add refusal steeringLlama-3.1-8B-Instruct · BBQ Race bias scorebias score: -1.15 -> -0.48No clear effect
add refusal steeringLlama-3.1-8B-Instruct · BBQ Appearance bias scorebias score: 15.51 -> 15.13No clear effect
vary refusal direction variant (10 sampled directions) (ablation)refusal steering · SBBJaccard similarity of responses: minimum 0.75No clear effect
add adversarial-suffix jailbreak instead of refusal steeringLlama-2-7B, Gemma-7B, Qwen-7B · SBBDPD consistency across runsHurt

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy