Turing AgentAgentic Formal Verification for Mathematical Research
Our vision
-
Fast
Verify at AI speed.
-
Faithful
Prove the claim as stated.
-
For TCS
Formal foundations for TCS.
OpenAI’s Unfinished Lean Proofs, Completed by Turing Agent
In October 2026, OpenAI released 722 mathematical manuscripts authored by frontier models. Of 121 theoretical computer science problems, 22 formalizations remained unfinished. Highlighted here are 3 of these problems with Lean-kernel-checked proofs.
How the harness works
Fix a faithful statement
A theorem-formalization agent writes the Lean statement from Mathlib and CSLib definitions, then reviews domains, quantifiers and assumptions against the source. Once checks pass, the statement is locked.
Grow a proof DAG in parallel
Proof agents split goals into claims only after Lean confirms the children imply the parent, prove independent branches concurrently, and keep an experience pool of what worked and what failed.
Verify independently
An independent Lean kernel rechecks the assembled proof with an axiom audit, and a review agent checks that the proof still follows the original argument.
What's new
Citation
If you use Turing Agent or the results on this page, please cite:
@techreport{turingagent2026,
title = {Turing Agent: Agentic Formal Verification
for Mathematical Research},
author = {{Turing Agent Team}},
institution = {Shanghai Artificial Intelligence Laboratory},
year = {2026}
}