Astra becomes first public model to solve any curated hard Erdős problems with Lean proofs
Jsevillamol · x · 2026-09-04
Greg Burnham proposes a new AI math benchmarking idea: curate a set of roughly Erdős-hard problems, require Lean formal proofs, and run all of them with a substantial compute budget. Astra turns out to be the first public model scoring above 0% — officially 2/68 solved, plus 3 more in ad hoc runs. Using Lean proofs as an anti-cheating mechanism could shape the next generation of hard math evals.
Related event: Epoch AI Launches FrontierMath Erdős; GPT-6 Astra First to Break Zero(10 posts)→
More from Models
- Opus 5.5 tops OpenRouter spending in 20 of 29 task categories — mmurph · 2026-10-21
- Meta team builds open-source 30B LLM from scratch in months, hails 'second era' — TimDarcet · 2026-10-21
- DFlash drafter for Muse Glimmer-30B: 5 layers, blockwise 16-token spec decoding — TimDarcet · 2026-10-21
- EvoPress quantization shrinks 30B model from 56GB to 19GB at near-bf16 accuracy — TimDarcet · 2026-10-21
- Radar chart: Muse Glimmer-30B benchmarked against Gemma4-31B and Qwen3.6-27B — TimDarcet · 2026-10-21
- Privacy evals: Muse Glimmer-30B leaks less than Qwen3.6-27B on CIMemories — TimDarcet · 2026-10-21