Autonomous agents discover new math theorems in open-world environment
progenitor414 · reddit · 2026-08-30
A paper introduces "Station," an open-world multi-agent environment where AI agents autonomously collaborate on mathematical research without a central coordinator.
Key Results:
- Achieved novel results on 5 out of 12 construction problems from AlphaEvolve and 2 case studies.
- Discoveries include a new infinite family of finite-field Kakeya sets, new exact 604-point kissing configurations in dimension 11, and improved bounds for discretized Kakeya and Erdős's minimum-overlap problems.
- Agents produced not just numerical constructions but also theorems and analyses explaining the mechanisms.
- All raw dialogues, proofs, and verification code are released, demonstrating the potential of AI in autonomous scientific discovery.
More from Research
- A beginner-friendly introduction to causal inference kicks off, Part 1 — Nasereliver · 2026-10-16
- FBI method trains for Pass@K with one sample per update at 12.3% of the cost — EmtiyazKhan · 2026-10-16
- Berkeley researchers sample humanoid robot policies directly from weight space with diffusion — berkeley_ai · 2026-10-16
- OneWM: unified causal training shows world and action modeling improve each other — IAAI233 · 2026-10-16
- Baidu's Virtual Head Attention tames the KV cache, FLOPs, quality triangle in long-context LLMs — baidu · 2026-10-16
- 5.8M-param Speech Event Detector adds laughs, breaths and背景 sounds to Whisper transcripts — Ayush Kumar Bar · 2026-10-16