ShallowDefend: Harnessing Shallow Safety Alignment for Multi-Turn Defender
ShallowDefend makes a model commit to an explicit safety decision before replying, then trains that decision with reinforcement learning. The authors report stronger multi-turn safety across 14 benchmarks.
ShallowDefend uses an **Assess → Decide → Respond** sequence: the model evaluates intent, emits a parseable safety decision, then generates a reply conditioned on that decision. The approach was tested across **two model families** and **14 benchmarks** spanning safety, utility, and helpfulness.
For systems handling multi-turn requests, make the safety decision an explicit step that can be inspected and evaluated, rather than relying only on the final response. The paper describes both rule-based self-verification and an LLM-judge reward option for training that decision.
ShallowDefend uses an **Assess → Decide → Respond** sequence: the model evaluates intent, emits a parseable safety decision, then generates a reply conditioned on that decision. The approach was tested across **two model families** and **14 benchmarks** spanning safety, utility, and helpfulness. For systems handling multi-turn requests, make the safety decision an explicit step that can be inspected and evaluated, rather than relying only on the final response. The paper describes both rule-based self-verification and an LLM-judge reward option for training that decision. The authors report stronger multi-turn safety while preserving utility, with a modest helpfulness cost. The results are from the paper’s controlled comparisons and do not establish performance across other models or deployment settings.
ShallowDefend makes safety a visible, separately evaluated decision before response generation, with results across two model families and 14 benchmarks. This gives a multi-turn alternative to conditional safety-adapter routing; the reported helpfulness cost also keeps utility trade-offs in view.