Today's AI scientist systems judge a whole proposal novelty, which is hard to … · @SuJinyan6 · TwKit

@SuJinyan6

Today's AI scientist systems judge a whole proposal novelty, which is hard to judge even for expert reviewers. Most papers are not novel in every single contribution, but incremental, so the authors propose judging at the level of individual claims instead. The current AI scientist system judge might kill 96% of the ideas humans went on to publish. Instead, they propose to use a claim-level profile that tells which contribution is covered by which paper and to what degree, and let a judge weigh that profile. This way, the novety checker can keep incremental ideas that build on prior work, and tells the generator exactly which claim to change. It could even serve as a training reward for AI-scientist agents. Glad to see more and more research about "taste".

曝光 556 · 评论 0 · 点赞 11 · 书签 3