Researchers train a 27B model to rank which proofs are worth keeping

single source· 1 articles · confidence: medium · first seen 2026-09-22 20:00 UTC

What this means for you

Nothing to act on yet. This is a preprint: the abstract names no released weights, code or dataset, and the interestingness score is the authors' own definition checked against their own utility proxy. If you curate a formal library, the ranking signal is worth testing; there is nothing to run.

An arXiv preprint dated 22 September 2026 proposes scoring a mathematical theorem by the ratio of its proof length to its statement length, reporting that this correlates strongly with downstream utility. Using proof difficulty conditioned on a set of premises, the authors train a 27B model that predicts that difficulty more accurately than frontier general-purpose models. Ranking generated theorems by the metric cut substantial or full overlap with Mathlib, a machine-checked library of formal mathematics, from 91.9% to 30.6%. The system generates candidates, selects the highest-scoring, and iteratively extends the library.

Key facts

  • ·The paper defines a theorem's intrinsic interestingness as the ratio of its proof length to its statement length. source
  • ·The authors trained a 27B-parameter model to predict proof difficulty given a set of premises, reporting it outperforms frontier general-purpose models at that task. source
  • ·Optimising for the metric reduced substantial or full overlap with Mathlib from 91.9% to 30.6%. source
  • ·The system generates candidate theorems, ranks them by the metric, and adds selected results to a self-expanding formal library. source
  • ·The preprint is dated 22 September 2026; the abstract gives no evaluation date or harness for the reported figures and reports no released model, code or dataset. source

What the sources say

Sources

The original reporting. Follow these — they did the work.

← the wire