AI Paper Atlas

Research map / Large language models / Reasoning and chain of thought

SafeCompass: Dynamic Chain-of-Thought Steering via Inference-Time Safety Signals

Zeyang Zhang, HAOTIAN XU, Linbao Li, Qi Sun, Xuebo Liu, Yu Li, Cheng Zhuo

ICML 2026 · Large language models › Reasoning and chain of thought · approach: Chain-of-thought training and analysis

Read the paperCheck your idea against 137,217 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes SafeCompass, an inference-time method that dynamically steers chain-of-thought reasoning of large reasoning models using safety signals to improve safety against adversarial prompts.

Problem

Large reasoning models' explicit chain-of-thought can be manipulated by adversarial prompts; existing inference-time CoT interventions use static heuristics that ignore the dynamic nature of reasoning, creating a safety-utility trade-off.

Approach

Replaces static CoT intervention heuristics with dynamic steering driven by inference-time safety signals, adjusting the reasoning trajectory as it unfolds.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
insert safety guidance at the beginning of CoT instead of noneQwen3-8B · 5 attacks avg ASR / Or-BenchASR and over-refusal: ASR 0.563 -> 0.125; Or-Bench 0.01 -> 0.68Mixed
insert safety guidance mid-CoT (200 tokens) instead of at the start (ablation)Qwen3-8B · 5 attacks avg ASRASR: 0.125 -> 0.395Hurt
add SafeCompass dynamic interventionQwen3-8B · 9 jailbreaks avg ASRattack success rate: 0.520 -> 0.020Improved
add SafeCompass dynamic interventionQwen3-8B · Or-Bench / GSM8K / MMLU-Proover-refusal and accuracy: Or-Bench 0.01 -> 0.01; GSM8K 0.915 -> 0.910; MMLU-Pro 0.719 -> 0.719No clear effect
intervene at low safety-score positions rather than high-score positions (ablation)SafeCompass on Qwen3-8B · 9 jailbreaks avg ASRASR: 41.56% -> 2.0%Improved
keep the control branch when computing safety direction (ablation)SafeCompass · 4 jailbreaks avg ASR, Qwen3-8BASR: 0.395 -> 0.020Improved
keep neutral control tokens (instead of S_control = S_t) (ablation)SafeCompass · Or-Bench, Qwen3-8Bover-refusal: 0.400 -> 0.001Improved
use comma instead of period in safety sentence (ablation)SafeCompass · 4 jailbreaks avg ASR, Qwen3-8BASR: 0.153 -> 0.020Improved
change content of safety sentence, keeping trailing comma (ablation)SafeCompass · 4 jailbreaks, Or-Bench, GSM8K, MMLU-ProASR and utility: Qwen3 JB 0.020 -> 0.010No clear effect
use dynamic on-the-fly safety direction instead of fixed Jailbreak Antidote direction (ablation)SafeCompass · Qwen3-8B, JB and Or-BenchASR and over-refusal: JB 0.088 -> 0.020; OR 0.280 -> 0.001Improved
allow repeated interventions instead of a single one per generation (ablation)SafeCompass · avg ASR across models and attacksASR: obfuscation attacks ASR +19.8% under single-intervention limitImproved
relax intervention (lower alpha or larger n) (ablation)SafeCompass · Qwen3-8B, DeepSeek-8BASR: slight rise in ASRHurt

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

Experimental setup

Baselines: static CoT intervention heuristics

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