AI Prover–Verifier Pipeline Reports Progress on Nine Open Problems
Binghui Peng, Runzhou Tao, Steven Wang, and Hantao Yu report that a GPT-5.5 Pro and Claude Opus 4.8 prover–verifier workflow produced results on nine open problems across learning theory, algorithms, and commutative algebra, with one result explicitly partial and no independent review of the full collection established by the retained sources.
Hantao Yu and Binghui Peng report joint work with Runzhou Tao and Steven Wang using a simple AI prover–verifier pipeline on nine open problems: four from the COLT open-problem track, one arising from a FOCS 2023 paper, and four from commutative ring theory.
The project identifies GPT-5.5 Pro as the proof generator and Claude Opus 4.8 as the verifier. Its reported targets include shuffled-SGD inequalities, robust conditional-probability estimation, measured-output quantum-circuit learning, unweighted data selection for linear regression, adversarial robustness of online leverage-score sampling, and four problems from Cahen–Fontana–Frisch–Glaz; the measured-output quantum-circuit result is explicitly labeled partial. The authors also report an agentic pipeline that formalizes the four commutative-algebra solutions in Lean 4.
The public repository collects problem write-ups, pipeline code, and Lean formalizations, but the retained materials do not establish independent review of the full collection. Accordingly, this item records the authors' claims as unverified progress and does not treat all nine entries as complete solutions.
Sources: Hantao Yu's announcement · Binghui Peng's announcement thread · Binghui Peng's detailed project announcement
Related Materials: Pipeline Math repository