GPT-6 Astra helps crack first Major Advance problem on FrontierMath
Epoch AI announced that another open problem in its FrontierMath: Open Problems benchmark has been solved, marking the benchmark's first 'Major Advance' result. The solvers were mathematicians Becker, Greger and Dominik Peters, who elicited the solution through long interactive sessions with GPT-6 Astra. According to 新智元, the researchers are affiliated with TU Munich (Patrick Becker) and Oxford (Matthias Greger), and the result can be summarized as proving that an absolutely fair committee must exist.
Confirmed
- The problem was an open question in voting theory posed in 2017 (unsolved for roughly 8-9 years), related to the line of research in 'The Core in Approval-Based…'.
- The model did not solve it autonomously: per kimmonismus and haider1, researchers relied on key ideas and proofs supplied by GPT-6 Astra, which even invented a new voting rule, with human mathematicians completing the proof.
- This is the first problem on FrontierMath rated at 'Major Advance' level. FrontierMath, designed by Epoch AI, contains research-level problems previously far beyond model capabilities.
Why it matters
- Multiple posters (kimmonismus, haider1) view this as AI moving from benchmark-gaming to real mathematical research, providing substantive research ideas rather than reproducing known answers.
- haider1 notes GPT-6 Astra is the forward-looking successor to Astra per Sam Altman, adding signal value to the achievement.
2026-09-19 ~ 2026-09-20 · 8 related posts
Primary sources
- AI solves FrontierMath's first 'major advance' tier problem — ResultBackground2450 · 2026-09-19
- FrontierMath open problem solved via interactive session with GPT-6 Astra — RexDouglass · 2026-09-19
- GPT-6 Astra reportedly solves FrontierMath's first Major Advance problem, devising a novel voting rule — haider1 · 2026-09-19
- [source] GPT-6 Astra Helps Solve 2017 Voting-Theory Problem, FrontierMath's First Major Advance — kimmonismus · 2026-09-19
4 near-duplicate retellings: Jsevillamol · kimmonismus · Jsevillamol · 新智元