Back to news
researchAnthropicOpenAI2026-10-30

Research agent ClawsGO Science solves 50-year-old global attractor conjecture in chemistry in 11 days

The BitMiracle AI Lab team recently released a 23-page academic paper with synchronously open-sourced source code, claiming its research agent ClawsGO Science took only 11 days to crack the half-century-old global attractor conjecture in chemical kinetics and delivered fully machine-verified compila

The BitMiracle AI Lab team recently released a 23-page academic paper with synchronously open-sourced source code, claiming its research agent ClawsGO Science took only 11 days to crack the half-century-old global attractor conjecture in chemical kinetics and delivered fully machine-verified compilation with zero errors. Chemist Fritz Horn proposed the global attractor conjecture in 1974, asking whether in reversible reactions satisfying detailed balance, with total mass conserved, the system always converges to the same unique equilibrium regardless of extreme initial ratios, or whether some species can be exhausted in edge cases leaving the reaction stuck at a false equilibrium. For half a century scholars have struggled with boundary false equilibria and complex nonlinear dynamics without a fully rigorous proof. The breakthrough came from a multi-agent structure: Claude Opus 5.5 as strategic brain setting research taste and technical route, decomposing the problem and correcting course at key moments; a strike team of 136 concurrent GPT 6 Astra agents writing proofs, adapting 71 formalization modules and evolving code in batches; and a Lean 4 formalization kernel as strict examiner, with every proof line compiled by the machine kernel so a single type error fails. On day six the work stalled at the convergence proof, with the textbook LaSalle invariance principle initially chosen, but Astra agents reported the path depended on heavy formalization of initial-value continuity and risked a dead end, so the Opus 5.5 decision brain switched to an elementary geometric energy-descent strategy, and the whole line compiled successfully the next day.

ClawsGO Science全局吸引子猜想多智能体Lean 4形式化验证自主科研