SafeSpec: Fast and Safe LLM via Dynamic Reflective Sampling

arXiv·medium signal

Reconciles speculative-decoding speedups with safety by attaching a lightweight safety head that judges semantic validity and safety in one forward pass during verification, applying rollback and safety-guided resampling instead of halting when unsafe output appears. On Qwen3-32B it cuts attack success rates by 15% while keeping a 2.06x inference speedup on benign inputs. Useful for teams who want both fast and guarded generation.