Research map / Large language models / Agents, code and math
ExVerus: Verus Proof Repair via Counterexample Reasoning
Jun Yang, Yuechun Sun, Yi Wu, Rodrigo Caridad, Yongwei Yuan, Jianan Yao, Shan Lu, Kexin Pei
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 |
|---|---|---|---|
| Guide repair with source-level counterexamples (ExVerus) instead of AutoVerus handcrafted repair prompts | Verus proof repair · VerusBench | success rate (%): 24.7 -> 71.9 (DeepSeek-V3.1) | Improved |
| Add counterexample-guided mutation and validation over direct fix from counterexamples (ablation) | ExVerus · VerusBench (DeepSeek-V3.1) | success rate (%): 64.4 -> 71.9 | Improved |
| Add counterexample-guided mutation and validation (ablation) | ExVerus · ObfsBench (DeepSeek-V3.1) | success rate (%): 65.4 -> 81.6 | Improved |
| Add mutation and validation (ablation) | ExVerus · DafnyBench (o4-mini) | success rate (%): 85.1 -> 95.5 | Improved |
| Add mutation and validation (ablation) | ExVerus · DafnyBench (GPT-4o) | success rate (%): 89.6 -> 88.1 | Hurt |
| Use 10 counterexamples instead of one (ablation) | ExVerus · InvariantInjectBench | tasks proved (of 187): 100 -> 106 | Improved |
| Mutant blocks at least one validated counterexample (ablation) | ExVerus mutant ranking · InvariantInjectBench | mutant verification rate: 38.55% -> 64.49% | Improved |
| Use ExVerus under code obfuscation vs AutoVerus | Verus proof repair · ObfsBench | success rate (%): above 73% vs below 40% | Improved |
| Use ExVerus instead of AutoVerus on cost | Verus proof repair · DeepSeek-V3.1, per task | cost (USD): 0.17 -> 0.04 | Improved |
| Use a stronger LLM (Sonnet-4.5) with ExVerus on easy benchmarks | ExVerus · DafnyBench | success rate (%): 95.5 -> 95.5 | No clear effect |
| ExVerus on LCBench with weaker models | ExVerus · LCBench (DeepSeek-V3.1) | success rate (%): 10.7 -> 10.7 | No clear effect |
Closest papers by meaning
- APOLLO: Automated LLM and Lean Collaboration for Advanced Formal Reasoning (NeurIPS 2025 · similarity 0.88)
- AlphaVerus: Bootstrapping Formally Verified Code Generation through Self-Improving Translation and Treefinement (ICML 2025 · similarity 0.86)
- Variation in Verification: Understanding Verification Dynamics in Large Language Models (ICLR 2026 · similarity 0.86)
- Lemur: Integrating Large Language Models in Automated Program Verification (ICLR 2024 · similarity 0.86)
- FaithSieve: Fine-Grained Evaluation of Math Proofs with Faithful Formal Evidence (NeurIPS 2026 · similarity 0.86)
- FVEL: Interactive Formal Verification Environment with Large Language Models via Theorem Proving (NeurIPS 2024 · 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.
