Propose, Solve, Verify: Self-Play Through Formal Verification

Fuente: arXiv
Guardado en:
Detalles Bibliográficos
Autores principales: Wilf, Alex, Aggarwal, Pranjal, Parno, Bryan, Fried, Daniel, Morency, Louis-Philippe, Liang, Paul Pu, Welleck, Sean
Formato: Preprint
Publicado: 2025
Materias:
Acceso en línea:
Etiquetas: Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
_version_ 1866915688203943936
author Wilf, Alex
Aggarwal, Pranjal
Parno, Bryan
Fried, Daniel
Morency, Louis-Philippe
Liang, Paul Pu
Welleck, Sean
author_facet Wilf, Alex
Aggarwal, Pranjal
Parno, Bryan
Fried, Daniel
Morency, Louis-Philippe
Liang, Paul Pu
Welleck, Sean
contents Training models through self-play alone (without any human data) has been a longstanding goal in AI, but its effectiveness for training large language models remains unclear, particularly in code generation where rewards based on unit tests are brittle and prone to error propagation. We study self-play in the verified code generation setting, where formal verification provides reliable correctness signals. We introduce Propose, Solve, Verify (PSV) a simple self-play framework where formal verification signals are used to create a proposer capable of generating challenging synthetic problems and a solver trained via expert iteration. We use PSV to train PSV-Verus, which across three benchmarks improves pass@1 by up to 9.6x over inference-only and expert-iteration baselines. We show that performance scales with the number of generated questions and training iterations, and through ablations identify formal verification and difficulty-aware proposal as essential ingredients for successful self-play.
format Preprint
id arxiv_https___arxiv_org_abs_2512_18160
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Propose, Solve, Verify: Self-Play Through Formal Verification
Wilf, Alex
Aggarwal, Pranjal
Parno, Bryan
Fried, Daniel
Morency, Louis-Philippe
Liang, Paul Pu
Welleck, Sean
Artificial Intelligence
Training models through self-play alone (without any human data) has been a longstanding goal in AI, but its effectiveness for training large language models remains unclear, particularly in code generation where rewards based on unit tests are brittle and prone to error propagation. We study self-play in the verified code generation setting, where formal verification provides reliable correctness signals. We introduce Propose, Solve, Verify (PSV) a simple self-play framework where formal verification signals are used to create a proposer capable of generating challenging synthetic problems and a solver trained via expert iteration. We use PSV to train PSV-Verus, which across three benchmarks improves pass@1 by up to 9.6x over inference-only and expert-iteration baselines. We show that performance scales with the number of generated questions and training iterations, and through ablations identify formal verification and difficulty-aware proposal as essential ingredients for successful self-play.
title Propose, Solve, Verify: Self-Play Through Formal Verification
topic Artificial Intelligence
url https://arxiv.org/abs/2512.18160