Together AI launched EinsteinArena, an open platform where AI agents work together on unsolved science problems. The agents post to discussion threads, submit solutions through a live API, and build on each other's partial results, all scored by deterministic verifiers in sandboxed environments.
The headline result: agents collectively improved the lower bound for the Kissing Number problem in dimension 11 from 593 to 604. That previous bound of 593 was set by Google DeepMind's AlphaEvolve in mid-2025, and before that, the number had been stuck at 592 for two decades. The jump to 604 is large by the standards of this problem. According to Together AI's blog post, the breakthrough started on April 8 when one agent submitted a promising but invalid construction with overlapping spheres. Other agents then spent 48 hours refining it, trading the top leaderboard spot in real time. The verified 604-sphere solution landed on April 11.
No single agent solved it alone. The chain of refinements required techniques like LSQR optimization and integer snapping to push overlap error from 1e-13 down to 1e-50. Together AI had to upgrade the verifier overnight because standard floating-point arithmetic couldn't handle the required precision.
Beyond the Kissing Number, agents on the platform have produced 11 new best-known results across problems including the Erdős minimum overlap problem, autocorrelation inequalities, circle packing, and the Tammes problem. The platform launched on March 19, 2026, and the full list of results is tracked in a public GitHub repository. Together AI says it plans to expand beyond math into computational biology and proof-based problems.
Bottom Line
AI agents collaborating on EinsteinArena improved the 11-dimensional Kissing Number lower bound from 593 to 604, surpassing AlphaEvolve's 2025 result in under 48 hours of collective refinement.
Quick Facts
- Kissing Number dimension 11 lower bound: 593 → 604
- Previous best (593) set by Google DeepMind's AlphaEvolve in mid-2025
- 11 new best-known results across open math problems
- Platform launched March 19, 2026
- Breakthrough construction submitted April 8; verified April 11, 2026
