AI novelty filter screened out 96% of papers published later

A new paper from the University of Illinois Urbana-Champaign stress-tests the novelty-checking step in AI research pipelines. The paper is titled “Do AI Scientists Know What's New?” and its authors include Miri Liu, Xiayu Zhao, Encheng Xie, and ChengXiang Zhai; arXiv ID 2610.36511.

The background is straightforward. After an AI scientist pipeline generates a batch of research ideas, it runs a novelty check to decide which ideas deserve compute. The problem is that such checks usually treat an entire proposal as a single unit when judging whether it is novel. Real papers rarely grow from scratch; most build on existing contributions. Judging the whole proposal at once makes it easy to treat “has old components” as “has no novelty.”

In the tweet, Rohan Paul gives this comparison: an already deployed novelty checker caught all 380 disguised copies while rejecting 96% of ideas that human researchers later published. ClaimNov works differently: it breaks a proposal into several claims, matches them one by one against claims in prior papers, and then gives the claim-level evidence to a judge. As a result, it catches about 97% of disguised copies and rejects only 1.5% of ideas that were later published. The figure caption gives more detailed numbers: ClaimNov correctly rejected 367 disguised papers, identified the source in 366 of them, and falsely rejected 1.5% of 400 ideas later implemented by humans.

The 96% figure is more worth looking at than the 380 papers. A filter that rejects everything can also catch all disguised copies. In a reply, Lakshay said that rejecting everything would also catch all 380 papers, and that 96% is the only number that really matters. Iris晴柔 added another side: rejected ideas never get run, so the checker never has a chance to know it judged wrongly. This feedback loop is very narrow, and the errors are hidden.

Elev’s comment was more cutting: a checker that rejects 96% of later-published papers has completely learned to be Reviewer 2. The joke points to an evaluation-unit problem. Judging novelty at the level of the whole proposal amounts to requiring that an idea carry no trace of the old. But in reality, good ideas are often recombinations of old results. A commenter added that this checker, which rejected 96%, came from the study by Si et al. that recruited more than 100 NLP researchers; that study also concluded that “AI’s ideas are more novel than human experts’.” According to the reproduction in this paper, after removing the original paper from the library, it still rejected 359 of the 380 disguised manuscripts. This addition points to something more fundamental: the novelty checker is judging “does it look like existing work,” not “is it worth doing.”

The paper’s engineering recommendations are very concrete. The pipeline should filter ideas, not score whole proposals and apply a blanket cutoff. Break the proposal into claims, evaluate novelty separately, and keep proposals that build on one known result. Then leave an appeal channel. A commenter noted in a reply that papers published later show that a “not novel” judgment needs to be appealable. For automated research pipelines, the novelty checker is more like a triage desk; don’t let it be the judge. The cost of mis-triage is that there are no later experiments to correct it. Suspicion should trigger an investigation, not terminate it outright. Marcus van der Erve also mentioned in a reply that unconventional ideas are easily dismissed by existing scientific expectations; the turning point is to have AI develop an idea into testable propositions and then test them against evidence.

ClaimNov is not a perfect filter; its 97% recall and 1.5% false-rejection rate still need to be viewed within a specific pipeline. But the direction it points in is clear: novelty judgment should move down from the whole proposal to the claim level, separating “has old components” from “makes no new contribution.” For people building AI scientist tools, this is more useful than tuning another binary classification threshold.