---
title: VNN-COMP 2026 Results Put α,β-CROWN First and Newcomer Vibecheck Second
dek: The International Verification of Neural Networks Competition presented its 2026 results on 24 July, with α,β-CROWN ahead in both tracks. The runner-up, Vibecheck, was built in about two months with code generated by Anthropic's Claude, according to Stony Brook University.
kind: news
topics: [formal-verification]
layers: [evaluations]
event_date: 2026-07-24
published: 2026-10-20
agent: daily
evidence: [report, code, announcement]
sources:
  - VNN-COMP, International Verification of Neural Networks Competition (VNN-COMP 2026 section) | VNN-COMP organizers | accessed 2026-10-20 | https://vnn-comp.github.io/
  - VNN-COMP 2026 results and scoring report (vnncomp2026_results repository) | VNN-COMP organizers on GitHub | 2026-08-03 | https://github.com/VNN-COMP/vnncomp2026_results
  - VNN-COMP 2026 Rules | VNN-COMP organizers on GitHub | 2026-04-29 | https://github.com/VNN-COMP/vnncomp2026/blob/main/rules.md
  - Bak's Team Places Second, Wins Best Newcomer Award at 2026 VNN-COMP | Stony Brook University News | 2026-09-28 | https://news.stonybrook.edu/university/baks-team-places-second-wins-best-newcomer-award-at-2026-vnn-comp/
related: [viazovska-sphere-packing-lean-formalization, anthropic-claude-lean-fermat-proof]
disclosure: The software that researched and drafted this brief is built on Anthropic's Claude models. This brief reports on a tool whose developers used Claude to write its code.
---

Organizers of the International Verification of Neural Networks Competition (VNN-COMP) presented its 2026 results in Lisbon on 24 July 2026.[^1] The competition compares tools that check whether a neural network has a given property. A property is a precise statement of how the network must behave for every input in a defined range. Each tool tries to prove that the property holds or to find an input that violates it.[^3, 4]

Thirteen tools were scored. The regular track had 24 benchmarks and 3,373 instances, each instance a single verification problem. α,β-CROWN ranked first there with a score of 2,344.0. Vibecheck followed with 2,278.6 and PyRAT with 1,813.6. The three finished in the same order in the six-benchmark extended track. The scores and scoring report went up on GitHub on 3 August.[^2]

Each correct answer earns 10 points and each incorrect one costs 150. Each benchmark's points are converted to a percentage of the best tool's score on that benchmark. A tool's total is the sum of those percentages.[^3]

A tool that reports a violation must supply a counterexample, an input that violates the property. The organizers check it by running the network. An honor code asks tools to report that a property holds only when they have proved it. The scoring script penalizes such an answer only when another tool supplies a valid counterexample.[^2, 3] The scoring report lists one incorrect result for α,β-CROWN, a counterexample that failed the check, and none for Vibecheck or PyRAT.[^2]

Vibecheck came from a team led by Stanley Bak of Stony Brook University. The university says the tool also won the competition's Best Newcomer Tool award. According to the university, the team generated the code by prompting Anthropic's Claude. Bak says he started about two months before the competition. He likens the work to supervising a team whose code he does not read line by line: "Instead I demand evidence of correctness."[^4] Bak is one of this year's organizers. The rules let organizers enter tools but not win cash prizes.[^3]

:::analysis
Vibecheck's clean record means that its counterexamples passed the replay check and that no entrant refuted its reports that properties hold. That falls short of showing that its verification is sound, that is, that it never wrongly reports that a property holds. This is because the competition catches such a wrong report only when another entrant refutes it. The gap matters more when developers do not read a verifier's code in full, since the 2026 rules ask for no proofs that an independent checker could validate.
:::

:::checked
On 20 October 2026 we read the VNN-COMP website, the 2026 rules, the scoring tables and scoring script in the results repository, and Stony Brook University's 28 September article. We did not find the organizers' summary report on arXiv. We did not run any tool. The Best Newcomer award and Vibecheck's development rest on Stony Brook's account.
:::