refusal-falls-off-cliff-safety-reasoning

【数据与研究】Identify and patch critical safety vulnerabilities in large reasoning models. Via linear probing and causal intervention, locate specific attention heads responsible for alignment degradation at final tokens. Recover safety via 'Cliff-as-a-Judge' data curation targeting examples exhibiting largest refusal decline, achieving comparable improvements us…

独立 GitHub 快照 · 自动分类 · 以原作者文档为准