Live Lean 4 benchmark tracks Riemann Hypothesis proof compression across models and agents
flngr · x · 2026-11-22
A live platform now tracks substantive mathematical and Lean 4 proof-engineering simplifications across the Riemann Hypothesis formalization, covering analytic number theory, Hironaka resolution, and Levent Alpöge's full 980.4M-character RH formalization.
- Uses an anti-golf normalized token metric on Core Pillars 1–5 (comments, whitespace, imports stripped) to measure compression without score-gaming
- Every breakthrough is verified under Lean 4.34.1 + Comparator (0 sorry, standard axioms)
- Includes a Model × Agent Harness benchmark grouping verified submissions by model and harness, with a 0–100 index averaged over the 5 pillars
- Key techniques come with paper-style LaTeX derivations and Mathlib theorem substitutions
It effectively turns frontier math formalization into a continuous public benchmark for models and agent harnesses.
More from Research
- Kalshi traders now put 54% odds on P vs NP being resolved before 2030 — ricklamers · 2026-11-22
- OpenAI proof dump shows first superlinear circuit lower bound, but progress stays slow — thomasahle · 2026-11-22
- One photo, 8 VLAs, only 3 correct: ERQA embodied-reasoning benchmark questioned — giffmana · 2026-11-22
- Fleuret: treat AI as oracles, not explainers — human-comprehensible math will be negligible — francoisfleuret · 2026-11-22
- Vovk paper shows probabilistically nonrandom objects have vanishing universal measure — _onionesque · 2026-11-22
- AI generates a Riemann proof, AI generates the rebuttal—and it scores 100% AI — zetalyrae · 2026-11-22