Turing Agent Machine-Checks Three Unresolved Theory Tasks
A team of 16 researchers led by Yang Chen posted a paper on arXiv on Nov. 16 describing Turing Agent, a system that checks research-level math proofs with formal verification. The authors say it completed the formal verification of three previously unresolved theoretical computer science tasks from OpenAI's October 2026 collection.
Formal verification means a computer checks every step of a proof, not just a human reader. The paper says AI is producing math proofs faster than people can review them. The authors ask how to turn that growing pile into knowledge that machines have confirmed.
Turing Agent is a harness, meaning software that coordinates AI models through a long job. According to the abstract, it combines agents that formalize theorems and agents that formalize proofs. It also breaks proofs into pieces, runs checks in parallel, keeps what it learned from earlier proofs, and reviews meaning selectively. The team tested it on theoretical computer science, a field they describe as demanding because of tangled dependencies and strict requirements on meaning.
The authors compared it with direct generation and with general-purpose coding tools, including Codex CLI and Claude Code. They report about 57% to 62% faithful formalization success on Formal TCS Bench. They also report 83% overall pass@1 on a theory-focused subset of RLMEval. Those numbers used GLM-5.3 and DeepSeek-V4-Pro, two entries you can compare in our list of AI models. The version built on DeepSeek finished the three open tasks. One of them needed about a hundred formally verified supporting claims.
The paper also says the process sometimes turned up simpler proofs and exposed dependencies the original proofs did not need. It can leave behind reusable math knowledge as well. Checking AI output is a wider theme, as in Amazon's BenchProofer for coding benchmarks.
Much is still unknown. The arXiv page shows a first version, and the source does not say whether outside reviewers have examined it. The results are the authors' own, and the evaluation covers only theoretical computer science. The abstract does not say how the system performs in other areas of math or what it costs to run.
Sources
Cite this
Turing Agent Machine-Checks Three Unresolved Theory Tasks, New Mind Museum, November 17, 2026. https://newmindmuseum.com/research/2026/11/turing-agent-formal-verification-math/
New Mind Museum. (2026, November 17). Turing Agent Machine-Checks Three Unresolved Theory Tasks. https://newmindmuseum.com/research/2026/11/turing-agent-formal-verification-math/
“Turing Agent Machine-Checks Three Unresolved Theory Tasks.” New Mind Museum, 17 Nov. 2026, https://newmindmuseum.com/research/2026/11/turing-agent-formal-verification-math/.
@misc{new-mind-museum-2026-turing-agent-machine-checks-three-unreso,
author = {{New Mind Museum}},
title = {Turing Agent Machine-Checks Three Unresolved Theory Tasks},
year = {2026},
url = {https://newmindmuseum.com/research/2026/11/turing-agent-formal-verification-math/}
}Free to quote and cite with a link to this page.