SHARD: Safe and Helpful Alignment via Self-Reframing Distillation

The quietly significant detail is the closed-loop nature of the training signal: SHARD never consults a larger or separately trained model to define what counts as safe, which means the alignment boundary is entirely a function of what the base model can already reason about when prompted to reframe. That's a meaningful architectural constraint, not just a cost optimization.

The EIBench paper covered the same day surfaces a useful contrast: where EIBench measures whether models sustain appropriate behavior across multi-turn emotional dynamics, SHARD targets the earlier problem of whether a model will engage at all rather than refuse. Both are attacking alignment failures, but from opposite ends of the over-refusal versus under-refusal spectrum. The stance-detection work on LLM-assisted interpretive coding is also relevant here, because it exposed how poorly zero-shot prompting handles theoretically loaded constructs. SHARD's reframing step is essentially a structured prompt intervention, and the same ambiguity problems that undermined zero-shot coding could affect how reliably a model reframes genuinely ambiguous sensitive prompts versus clearly benign ones.

Watch whether SHARD's helpfulness gains hold when tested against adversarial reframing, where a user deliberately constructs a prompt that looks benign after rewriting but isn't. If the technique degrades on red-team evaluations published in the next six months, the closed-loop assumption becomes its primary liability.