FAR shifts AI mathematics from solving one chosen problem to searching a whole research direction
Posted to arXiv on August 17, 2026, the FAR pipeline shows how AI can scan mathematical literature, surface open conjectures, attempt them at scale, and send only the strongest artifacts to expert review.
AI mathematics is starting to optimize which problems deserve attention
An arXiv paper posted on August 17, 2026 introduces Find, Attempt, and Recommend (FAR), a pipeline for AI-assisted mathematical discovery. The notable change is where the workflow begins. Instead of handing the system one theorem, conjecture, or benchmark problem, the human gives it a research direction. The system then searches the literature for candidate open problems, filters them for status and relevance, attempts them, and returns a smaller set of artifacts for mathematicians to inspect.
Why this matters for AI in mathematics
This tackles a real bottleneck in research workflows. As frontier models get better at attempting mathematics, the scarce resource is no longer only raw reasoning time. It is also expert attention: which problems are worth trying, which candidate results deserve checking, and how to avoid wasting review effort on weak or already-resolved claims. FAR treats that allocation problem as part of the mathematics pipeline itself.
What the paper reports
According to the arXiv abstract, the authors build a literature-to-review cascade inspired by search and recommender systems. In a combinatorics pilot, the pipeline starts from 5,245 papers, recovers 6,453 candidate conjectures or open problems, filters them to 4,717 apparently well-posed and still-open statements, and then surfaces 598 potential resolutions before selecting 77 items for author-team review. The abstract says this process produced several interesting discoveries, including work on problems associated with Davies--Jenssen--Perkins--Roberts, Erdos--Straus, Ikenmeyer--Pak--Panova, and Lund--Saraf--Wolf.
The deeper lesson
The important signal for MathsAI readers is that mathematical AI is becoming less like a chatbot waiting for a prompt and more like research infrastructure that helps decide where effort should go. That is a meaningful shift. In serious mathematics, choosing a fruitful problem can matter as much as solving it. A system that can recover source-grounded open questions, attach status evidence, and rank outputs for review may become more useful than one that only performs on prepackaged benchmarks.
What MathsAI readers should watch next
The open question is how well this approach generalizes beyond a single combinatorics pilot. If direction-level workflows scale to other areas of mathematics, they could change how researchers search the literature, triage conjectures, and organize collaboration with formal verification or symbolic tools. The main risk is obvious: a larger funnel can also produce more noise, so provenance, status checking, and expert review remain essential.