PapersBuilders09:38Chinese contentICA 认证机制:为 σ 加权后训练加上有效性闸门这篇 arXiv 论文讲了个挺反直觉的坑:训练时删掉的坏行为会换个形式回来。作者用配对干预把问题量化,再提出 ICA 闸门修好,做后训练的值得看看#SFT#RLHF#Tool-Integrated Reasoning#MATH-5001 个信源在谈EventaarXiv cs.AI@Leiyue Chen et al. (4)2 sourcesSourceRead laterReadWorth followingUsefulFollow SFT