RESEARCH · RESEARCH · #2106
arXiv paper: corrective feedback reduces discovery of specification counterexamples in Verus/Rust
The arXiv preprint (arXiv:2610.20676v1) shows that corrective correctness feedback can suppress discovery of programs that meet written contracts but violate intended behavior. On 12 Verus/Rust tasks across four model families under matched call budgets, independent no-feedback discovery more than doubles certified-witness yield versus IntentShield’s corrective-feedback loop (59.0% vs 27.1%); a phase-separated execution matches the aggregate yield without reusing responses, and preserving those witnesses yields stronger specification repair (Phase-Revised: 81.3% implementation success vs IS-Revised: 62.5%). The effect did not reproduce in available Dafny and Lean evaluations, suggesting an empirical boundary around the studied Verus/Rust workflow.
KEY POINTS
- The arXiv preprint (arXiv:2610.20676v1) shows that corrective correctness feedback can suppress discovery of programs that meet written contracts but violate intended behavior.
- On 12 Verus/Rust tasks across four model families under matched call budgets, independent no-feedback discovery more than doubles certified-witness yield versus IntentShield’s corrective-feedback loop (59.0% vs 27.1%); a phase-separated execution matches the aggregate yield without reusing responses, and preserving those witnesses yields stronger specification repair (Phase-Revised: 81.3% implementation success vs IS-Revised: 62.5%).
- The effect did not reproduce in available Dafny and Lean evaluations, suggesting an empirical boundary around the studied Verus/Rust workflow.
WHY IT MATTERS
This matters because common corrective-feedback loops can eliminate the very counterexamples needed for specification-completeness auditing and therefore weaken automated detection and repair of intended-behavior violations in model-driven verification pipelines.