Raising machine-checked security benchmarks to advance hash-based SNARKs through agentic collaboration

The Ethereum Foundation’s Formal Verification team, in partnership with Yukon and zkSecurity, has launched better.codes, an open autoresearch challenge that formalizes a Reed–Solomon proximity problem in Lean to test and improve the soundness bounds of koalaIRS12. Participants deploy AI agents to incrementally raise the proven lower bound toward a fixed 128-bit target, with each accepted submission verified by a Lean kernel and promoted on a public leaderboard. The challenge targets modern succinct non-interactive proof systems (SNARKs), which underpin zkrollups, zkVMs, and Ethereum’s post-quantum security roadmap, where current proven bounds fall short of conjectured security levels. By making submissions transparent and git-backed, the challenge enables solvers to build on prior work, avoid dead ends, and collectively advance the frontier of formally verified cryptographic proofs.

The initiative builds on the Proximity Prize research, which aims to prove or disprove conjectures about Reed–Solomon proximity gaps central to hash-based SNARKs. The koalaIRS12 problem, formalized in ArkLib (Lean 4’s library for verified arguments of knowledge), directly addresses these grand challenges outlined in Open Problems in List Decoding and Correlated Agreement by Gal Arnon, Dan Boneh, and Giacomo Fenzi. Earlier challenges like ecdsa.fail, zk.golf, and snark.fast demonstrated the effectiveness of this model in accelerating research frontiers, including quantum circuit design and post-quantum proving speed. better.codes extends this approach to cryptographic security benchmarks, fostering open collaboration where diverse AI setups compete against a common verified benchmark. Each submission’s theorem statement and proof are rigorously checked, with accepted results credited to solvers and their AI models, ensuring transparency and reproducibility in the research process.

Nearly all production hash-based SNARKs rely on proximity gaps and correlated agreement for Reed–Solomon codes, but the proven security bounds today remain below the conjectured levels required for 128-bit security guarantees. The better.codes challenge seeks to close this gap by incrementally raising the proven lower bound through open, verifiable, and public research. By pinning the theorem statement, parameter point, and verification harness, the challenge ensures that every submission adheres to a fixed benchmark, while allowing solvers to explore new lemmas, proof techniques, and impossibility results. Accepted proofs are promoted to the public repository, with all prior work preserved in git-backed diffs and submission notes, enabling solvers to skip redundant efforts and focus on advancing the frontier.

Participants can sign in with GitHub and clone the challenge repository to begin solving koalaIRS12, with the goal of proving a larger soundness lower bound scored in bits. The challenge is governed by program terms that outline eligibility, evaluation, awards, and payments, which may be adjusted as the challenge progresses. Today’s launch focuses on the soundness challenge for koalaIRS12, but the Ethereum Foundation plans to introduce additional challenges over time to further advance research in formally verified cryptographic systems. The model emphasizes parallel, independent setups working toward a common benchmark, ensuring that no single agentic configuration dominates and that progress accelerates through collective effort.