FAR cascade cuts 5,245 math papers to 77 for expert review

Start with 5,245 combinatorics papers. End with 77 items a human mathematician actually reads.

That's the funnel described in a new preprint from Zeyu Zheng, Shengtong Zhang, Jeremy Avigad, Prasad Tetali and Sean Welleck, which proposes a rearrangement of how humans and AI split the work of research mathematics. The authors describe the shift in one sentence. 'The human input is no longer a single problem selected in advance, but a research direction in which the experts have interest and expertise.' The system then searches a broad literature corpus for candidate problems in that direction.

The pipeline, called FAR (for Find, Attempt, Recommend), is what the paper calls a 'literature-to-review cascade that automates the search for suitable problems.' Applied to a combinatorics corpus, it recovers 6,453 candidate conjectures or open problems, filters them to 4,717 apparently well-posed and still-open ones, then surfaces 598 potential resolutions and passes 77 items to the author team. Among what came out, the authors point to 'results on conjectures and questions of Davies--Jenssen--Perkins--Roberts, Erdős--Straus, Ikenmeyer--Pak--Panova, and Lund--Saraf--Wolf.'

The abstract doesn't publish per-stage hit rates or say how many of those 77 items were genuinely new versus previously known to the field. It's one pilot, in one branch of mathematics. Two researchers we track have already circulated it.