Tag
论文提出 LadderEdit，通过将每个 LoRA 编辑适配器压缩为低秩草图，并在草图不满足重写、泛化和局部性要求时沿阶梯提升秩，实现 5.2 倍的存储节省，并在 50,000 次连续编辑上保持有效。
J-Wash is a tool that uses Anthropic's Jacobian lens to let users edit the behavior and identity of large language models by modifying token directions, then export the edited model as a standalone checkpoint without training or fine-tuning.
This paper introduces the FactualOpinionEditing with Evidence (FOE) benchmark to assess the ability to edit factual opinions in LLMs, and proposes a Self-Generated Evidence-Aligned method to improve opinion-evidence alignment.
Experimental study shows inserting first/second-person pronouns into headlines has mixed effects on human memorability and that LLMs often produce inaccurate or unnatural revisions.