ProofOptimizer: Training Language Models to Simplify Proofs without Human Demonstrations
2510.15700v1
cs.LG, cs.AI, cs.PL
2025-10-21
Авторы:
Alex Gu, Bartosz Piotrowski, Fabian Gloeckle, Kaiyu Yang, Aram H. Markosyan
Abstract
Neural theorem proving has advanced rapidly in the past year, reaching IMO
gold-medalist capabilities and producing formal proofs that span thousands of
lines. Although such proofs are mechanically verified by formal systems like
Lean, their excessive length renders them difficult for humans to comprehend
and limits their usefulness for mathematical insight. Proof simplification is
therefore a critical bottleneck. Yet, training data for this task is scarce,
and existing methods -- mainly agentic scaffolding with off-the-shelf LLMs --
struggle with the extremely long proofs generated by RL-trained provers. We
introduce ProofOptimizer, the first language model trained to simplify Lean
proofs without requiring additional human supervision. ProofOptimizer is
trained via expert iteration and reinforcement learning, using Lean to verify
simplifications and provide training signal. At inference time, it operates
within an iterative proof-shortening workflow, progressively reducing proof
length. Experiments show that ProofOptimizer substantially compresses proofs
generated by state-of-the-art RL-trained provers on standard benchmarks,
reducing proof length by 87% on miniF2F, 57% on PutnamBench, and 49% on
Seed-Prover's IMO 2025 proofs. Beyond conciseness, the simplified proofs check
faster in Lean and further improve downstream prover performance when reused as
training data for supervised finetuning.
Ссылки и действия
Дополнительные ресурсы: