Research map / Large language models / Agents, code and math
Propose, Solve, Verify: Self-Play Through Formal Verification
Alex Wilf, Pranjal Aggarwal, Bryan Parno, Daniel Fried, Louis-Philippe Morency, Paul Pu Liang, Sean Welleck
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| train only on formally verified solutions (add verification filter) (ablation) | PSV-Verus · Dafny2Verus / MBPP / HumanEval (transfer) | pass@1: 31.82 -> 65.63 on Dafny2Verus | Improved |
| add difficulty labels to proposer prompt (ablation) | PSV-Verus · all three benchmarks | average pass@k: 40.87 -> 43.31 | Improved |
| refresh proposer in-context examples each iteration (diversity) (ablation) | PSV-Verus · all three benchmarks | average pass@k: 38.71 -> 43.31 | Improved |
| filter proposed specs with a spec verifier (ablation) | PSV-Verus · spec filtering | inference compute: 2.1x increase in inference compute without it | No clear effect |
| scale questions per iteration from 4k to 32k (ablation) | PSV-Verus · Dafny2Verus (TTT) | pass@1: 59.5% -> 74.3% | Improved |
| scale questions per iteration from 1k to 32k (ablation) | PSV-Verus · MBPP (TTT) | pass@1: 22.3% -> 44.3% | Improved |
| scale questions per iteration from 1k to 32k (ablation) | PSV-Verus · HumanEval (TTT) | pass@1: 13.9% -> 26.3% | Improved |
| split a 1k question budget over 5 iterations instead of 1 (ablation) | PSV-Verus · Dafny2Verus (TTT) | pass@1: 21.0% -> 42.3% | Improved |
| split a 4k question budget over more iterations (ablation) | PSV-Verus · Dafny2Verus (TTT) | pass@1: 28.0% -> 46.9% | Improved |
| replace baselines with PSV self-play proposing new specs | RFT (expert iteration) · MBPP (TTT) | pass@1: 3.83 -> 36.78 | Improved |
| train on Dafny2Verus with PSV then test on held-out MBPP | PSV-Verus vs AlphaVerus · MBPP (transfer) | pass@1: 6.48 -> 25.25 | Improved |
| difficulty prompting for proposer targets | PSV proposer · proposed question pass rates | mean pass rate: easy 0.55, medium 0.45, hard 0.36 | Mixed |
Closest papers by meaning
- ReVeal: Self-Evolving Code Agents via Reliable Self-Verification (ICLR 2026 · similarity 0.89)
- VERSE: Verification-based Self-Play for Code Instructions (AAAI 2025 · similarity 0.88)
- Improving Rationality in the Reasoning Process of Language Models through Self-playing Game (ICML 2025 · similarity 0.87)
- Trust, But Verify: A Self-Verification Approach to Reinforcement Learning with Verifiable Rewards (NeurIPS 2025 · similarity 0.87)
- Learning to Self-Verify Makes Language Models Better Reasoners (ICML 2026 · similarity 0.86)
- On the self-verification limitations of large language models on reasoning and planning tasks (ICLR 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
