Paper recorded by Signals 4 on 2026-10-29 in cs.CL. Abstract reproduced from arXiv; link to the original below.
Published 2026-10-29 on arXiv · recorded by Signals 4 on 2026-10-30
Category: cs.CL · 自然语言处理 · first seen 2026-10-30
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. This brittleness has a mechanistic root: Alignment is shallow, with safety behavior concentrated in a model's first few output tokens, and existing defenses inherit the weakness, supervising the surface form