AI Signal 130
Boundary-aware self-distillation trains LLMs to refuse only harmful subsets of a topic
↗
A new paper proposes boundary-aware self-distillation to train LLMs to refuse only the harmful subset of a topic, and repairs coverage gaps in self-generated safety data.
Topic-level safety guards over-refuse benign prompts that contain dangerous-looking words, which breaks deployments like civics tutors that need to answer factual political questions. This method shapes refusal at the boundary between harmful and benign prompts within a topic, and also fixes the coverage gap where hard harmful prompts are silently dropped from training data. Engineers building safety-tuned models can use this to align refusal with deployment-specific policies.
Written by elseif from the cluster below · every claim links back to a sourceThe three things worth knowing
The paper formalizes narrow-boundary safety as refusing a harmful subset of a topic while answering the benign complement.
The standard self-generation pipeline drops 19.88% of prompts due to failed steering attempts, which the method repairs with escalating retry to 0.20% residual failures.
The method adds in-distribution benign data, including 11,955 surface-dangerous benign prompts across 18 semantic types, to reduce false refusals.
THE CLUSTER