Researchers introduced ProofCoder, an agentic system for generating software-verification proofs in the Rocq proof assistant, in a paper submitted on 10 November 2026. The authors say the system uses feedback during proof search to decide when to inspect a proof state, seek information, revise a proof fragment or backtrack.[1]
The paper describes ProofCoder as a coding-agent approach to proof synthesis. It contrasts this with systems in which a surrounding loop controls the retrieval of relevant premises, proof-state management and backtracking while a language model generates tactics. ProofCoder can work on individual tactics or whole proofs, according to the abstract.[1]
The authors report evaluating the system on 280 tasks from the AutoRocq Benchmark. They say their best-performing configurations raised proof success from 47% for AutoRocq to 60%, at approximately half the cost. They also report that removing continuous reasoning solved fewer lemmas at substantially higher cost, with more backtracking. These are the authors’ reported results; the saved source is an abstract, not the underlying evaluation details.[1]
Sources
Check or correct this page
Verification record
The digest below shows whether the text you are reading matches the published source. It does not show that the reporting is correct. The build computes it from this page's Markdown source and lists it in the site manifest.
e10f90fb1faeb220e6f53d9781e8b4e95fc6115b30216a1474df0d51f59c8c03