ShallowDefend puts an explicit safety decision before multi-turn replies
Original title: ShallowDefend: Harnessing Shallow Safety Alignment for Multi-Turn Defender
Assess, Decide, Respond is trained with reinforcement learning and led multi-turn safety on 14 benchmarks across two model families.
The paper argues that safety alignment is concentrated in a model's first few output tokens, so multi-turn harmful requests often succeed. ShallowDefend has the model assess intent, commit to a parseable safety decision, then reply conditioned on that decision, with the commitment trained by reinforcement learning at the payload turn. Rewards can be a self-verified rule or an LLM judge. Authors report the strongest multi-turn safety in a controlled comparison, with utility preserved and a modest drop in helpfulness.
Unconfirmed: Benchmark leadership and the helpfulness cost are the authors' reported results.
- The committed decision is a token span that can be checked by a rule.
- Evaluation covers safety, utility, and helpfulness across 14 benchmarks.
- Per-token KL analysis and probes treat the decision as a causal control point.
- Authors say helpfulness falls modestly while safety stays ahead.
From the original article
1 min read at arxiv.orgSafety alignment has enabled large language models to reliably refuse harmful requests directly, yet the same request distributed across a multi-turn conversation still succeeds far too often.
By Qiyao Ma · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.