AI Paper Atlas

Research map / Large language models / Language, safety and interpretability

SteeringSafety: Benchmarking Representation Steering in LLMs Across Safety Perspectives

Vincent Siu, Nicholas Crispino, David Park, Nathan W. Henry, Zhun Wang, Yang Liu, Dawn Song, Chenguang Wang

ICML 2026 · Large language models › Language, safety and interpretability · approach: Safety alignment of LLMs

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
remove the KL filter (NoKL) from direction selection (ablation)steering methods · refusal and hallucination perspectiveseffectiveness / entanglement: often more than doubles entanglementMixed
add CAST-style conditional gating instead of the KL filter (ablation)steering methods · refusal perspectiveeffectiveness / entanglement: matching NoKL effectiveness at entanglement near StandardImproved
add CAST-style conditional gating instead of the KL filter (ablation)steering methods · bias perspectiveeffectivenessHurt
add conditional steering to DIMDIM on Gemma-2-2B · FaithEval Inconsistenthallucination reduction: 54.5% reductionImproved
steer refusal with ACE or DIM instead of CAA, PCA or LATsteering methods · refusal perspectiveeffectiveness: exceed 50% on every model except Gemma-2-2BImproved
use PCA for extrinsic hallucination steering in Llama-3.1-8BPCA · hallucination perspectiveeffectiveness/entanglement ratio: 1.76Improved
use LAT for bias steeringLAT · bias perspectiveeffectiveness/entanglement ratio: 7.05 and 8.37Improved
steer refusal (suppress it) and measure social behaviorsrefusal steering · social behaviorsentanglement: degradation up to 76%Hurt
steer refusal (jailbreak) and measure commonsense moralityrefusal steering · commonsense moralityaccuracy: up to 26%Hurt
steer intrinsic hallucination and measure political viewshallucination steering · TwinViews political viewsleft-leaning agreement shift: 25% shift to the right to a 28% shift to the leftMixed
steer any perspective and measure reasoningsteering methods · reasoning (GPQA, ARC-C)entanglement: below 2% in every caseNo clear effect
scale Qwen-2.5 from 1.5B and 3B to 7B (ablation)steering methods · effectiveness-entanglement ratio rankingranking stabilityNo clear effect

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