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
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 changed | Where | Measured | Effect |
|---|---|---|---|
| insert safety guidance at the beginning of CoT instead of none | Qwen3-8B · 5 attacks avg ASR / Or-Bench | ASR and over-refusal: ASR 0.563 -> 0.125; Or-Bench 0.01 -> 0.68 | Mixed |
| insert safety guidance mid-CoT (200 tokens) instead of at the start (ablation) | Qwen3-8B · 5 attacks avg ASR | ASR: 0.125 -> 0.395 | Hurt |
| add SafeCompass dynamic intervention | Qwen3-8B · 9 jailbreaks avg ASR | attack success rate: 0.520 -> 0.020 | Improved |
| add SafeCompass dynamic intervention | Qwen3-8B · Or-Bench / GSM8K / MMLU-Pro | over-refusal and accuracy: Or-Bench 0.01 -> 0.01; GSM8K 0.915 -> 0.910; MMLU-Pro 0.719 -> 0.719 | No clear effect |
| intervene at low safety-score positions rather than high-score positions (ablation) | SafeCompass on Qwen3-8B · 9 jailbreaks avg ASR | ASR: 41.56% -> 2.0% | Improved |
| keep the control branch when computing safety direction (ablation) | SafeCompass · 4 jailbreaks avg ASR, Qwen3-8B | ASR: 0.395 -> 0.020 | Improved |
| keep neutral control tokens (instead of S_control = S_t) (ablation) | SafeCompass · Or-Bench, Qwen3-8B | over-refusal: 0.400 -> 0.001 | Improved |
| use comma instead of period in safety sentence (ablation) | SafeCompass · 4 jailbreaks avg ASR, Qwen3-8B | ASR: 0.153 -> 0.020 | Improved |
| change content of safety sentence, keeping trailing comma (ablation) | SafeCompass · 4 jailbreaks, Or-Bench, GSM8K, MMLU-Pro | ASR and utility: Qwen3 JB 0.020 -> 0.010 | No clear effect |
| use dynamic on-the-fly safety direction instead of fixed Jailbreak Antidote direction (ablation) | SafeCompass · Qwen3-8B, JB and Or-Bench | ASR and over-refusal: JB 0.088 -> 0.020; OR 0.280 -> 0.001 | Improved |
| allow repeated interventions instead of a single one per generation (ablation) | SafeCompass · avg ASR across models and attacks | ASR: obfuscation attacks ASR +19.8% under single-intervention limit | Improved |
| relax intervention (lower alpha or larger n) (ablation) | SafeCompass · Qwen3-8B, DeepSeek-8B | ASR: slight rise in ASR | Hurt |
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.
- Jailbreak Antidote — Jailbreak Antidote: Runtime Safety-Utility Balance via Sparse Representation Adjustment in Large Language Models (ICLR 2025) · better on 1 of 1 shared result
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
- AdvChain: Adversarial Chain-of-Thought Tuning for Robust Safety Alignment of Large Reasoning Models (ICLR 2026 · similarity 0.91)
- SAFEPATH: Preventing Harmful Reasoning in Chain-of-Thought via Early Alignment (NeurIPS 2025 · similarity 0.91)
- Towards Safe Reasoning in Large Reasoning Models via Corrective Intervention (ICLR 2026 · similarity 0.90)
- Internalizing Safety Understanding in Large Reasoning Models via Verification (ICML 2026 · similarity 0.90)
- Safety Recovery in Reasoning Models Is Only a Few Early Steering Steps Away (ICML 2026 · similarity 0.88)
- BadThink: Triggered Overthinking Attacks on Chain-of-Thought Reasoning in Large Language Models (AAAI 2026 · 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.
