Do AI Scientists Know What's New?

The paper introduces ClaimNov, a claim‑level novelty detection module designed to improve AI scientific discovery pipelines. By evaluating novelty at the proposal and claim levels using four datasets of 1,060 proposals, the authors show that existing checks either reject too many novel ideas or fail to isolate non‑novel parts of a proposal. ClaimNov achieves a 97% detection rate of disguised papers with only a 1.5% false‑alarm rate, significantly reducing false positives compared to prior methods and preserving proposals that build on established work.

Machine-generated by The Flow from the publisher's headline and feed description — not written or checked by a human. The full article lives at arXiv AI.

• llms