Large Reasoning Models (LRMs) demonstrate strong problem‑solving capabilities, yet their safety alignment often deteriorates when confronted with harmful queries. Existing safety‑enhancement techniques mainly rely on additional fine‑tuning or preference optimization and provide limited insight into the internal causes of safety failures. To address this gap, we conduct a token‑level positional analysis of refusal dynamics and uncover a localized vulnerability at the very start of reasoning, which we name Onset Refusal Collapse (ORC). Our experiments show that, under harmful prompts, the refusal‑related signal of LRMs drops sharply at the first generated token, leading to unsafe responses. Motivated by this observation, we introduce SafeToken, a lightweight inference‑time intervention that injects a learned continuous safety anchor precisely at the reasoning onset. By updating only a single token embedding, SafeToken effectively mitigates ORC, improves safety on harmful‑query benchmarks, and largely preserves reasoning performance. These findings suggest that safety failures in LRMs can stem from a transient breakdown at the critical transition from understanding to generation.
Review