The first token is critical: understanding the safety collapse of large reasoning models

Researchers found that for large reasoning models (LRMs) under harmful queries, refusal-related signals drop sharply at the first generated token, and proposed an inference-time intervention method, SafeToken, to mitigate this safety failure. The paper calls this phenomenon Onset Refusal Collapse (ORC), which is associated with the generation of unsafe responses. SafeToken injects a learned continuous safety anchor at the start of reasoning, updating only a single token embedding; the paper reports that it mitigates ORC, improving safety on harmful query benchmarks while largely preserving reasoning ability. The authors accordingly propose that LRM safety failures may stem from a momentary collapse in the critical transition from understanding to generation.