Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
To resolve the collateral damage of coarse, topic-level alignment, this work formulates narrow-boundary safety to distinguish nuanced harms from benign in-domain queries (e.g., refusing political manipulation while answering factual election questions). While standard refusal fine-tuning sharply escalates benign over-refusal—driving XSTest false positives from 2.00% to 74.00% on Qwen3-8B—training on matched boundary pairs and verified target-model responses mitigates this collapse. Consequently, this boundary-aware self-distillation slashes comply-side over-refusal on held-out pairs from 32.94% to 4.16% while maintaining an 87.72% refusal rate on targeted harmful prompts.
Alignment doesn't have to be a blunt instrument: training on contrastive, self-distilled boundary pairs slashes false-positive refusals from 32.94% to 4.16% without compromising safety margins.
Safety alignment is usually posed as a topic-level question: is this subject harmful? Deployments ask a narrower one. A civics tutor and a public-sector assistant may share a base model yet need different boundaries inside the same topic, refusing targeted political manipulation while still answering factual questions about the same election. We formulate this as narrow-boundary safety and introduce an offline self-generated framework combining controlled topic generation, coverage repair, in-distribution compensation data, and harmful-benign pairs for training and evaluation. Single-shot generation leaves 19.88% of prompts without accepted refusal traces, whereas escalating retries leave 0.20%. On political persuasion with Qwen3-8B, training on refusal data completed through Escalate increases target-domain refusal from 9.47% to 84.75% and reduces the mean unsafe-response rate across three broader harmfulness benchmarks from 26.26% to 0.14%, but increases XSTest over-refusal from 2.00% to 74.00%. In a separate matched comparison, replacing external responses with verified target-model responses reduces over-refusal from 15.20% to 5.20%. Boundary-pair data reduces comply-side over-refusal on held-out pairs from 32.94% to 4.16%, while harmful-side refusal decreases only from 91.88% to 87.72%. These results show that data composition controls the safety and usability trade-off, and that safety alignment should be evaluated on both sides of the intended refusal boundary.