ShallowDefend: Harnessing Shallow Safety Alignment for Multi-Turn Defender
arXiv cs.CL · 30 Oct 2026, 03:56 AEST
ShallowDefend introduces a framework to enhance safety alignment in multi-turn conversations with large language models. By focusing on a short span of output tokens, the approach aims to improve the model's ability to refuse harmful requests effectively.
Key facts
- Safety 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.
- We propose Assess → Decide → Respond, a framework in which the defender assesses the user's intent, commits to an explicit, parseable safety decision, and only then generates a reply conditioned on it.
- Across two model families and fourteen benchmarks spanning safety, utility, and helpfulness, our defenders achieve the strongest multi-turn safety in a controlled comparison, maintaining their lead in safety while preserving utility at a modest cost to helpfulness.
Read aloud
ShallowDefend introduces a framework to enhance safety alignment in multi-turn conversations with large language models. By focusing on a short span of output tokens, the approach aims to improve the model's ability to refuse harmful requests effectively. Safety 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. We propose Assess → Decide → Respond, a framework in which the defender assesses the user's intent, commits to an explicit, parseable safety decision, and only then generates a reply conditioned on it. Across two model families and fourteen benchmarks spanning safety, utility, and helpfulness, our defenders achieve the strongest multi-turn safety in a controlled comparison, maintaining their lead in safety while preserving utility at a modest cost to helpfulness.
Sources in this cluster
- ShallowDefend: Harnessing Shallow Safety Alignment for Multi-Turn Defender
arXiv cs.CL · 30 Oct 2026, 03:56 AEST