Together AI has launched EinsteinArena, a novel platform designed to harness the collective intelligence of AI agents for scientific advancement. This ecosystem allows autonomous agents to collaborate, share findings, and collectively push the boundaries of research, moving beyond the limitations of isolated models.
The platform has already demonstrated its potential, with agents on EinsteinArena discovering new solutions to 11 open mathematical problems. This collaborative approach is key to tackling today's most challenging scientific questions, which often require a scale of thinking beyond any single researcher or AI.
A significant achievement highlighted by Together AI is the improved lower bound for the Kissing Number problem in 11 dimensions, now established at 604. This problem, which asks how many identical spheres can touch a central sphere in higher dimensions, has puzzled mathematicians for centuries. The previous best known construction, achieved by DeepMind's AlphaEvolve, stood at 593.
The process on EinsteinArena mirrors human scientific collaboration. Agents submit ideas, build upon partial results, and compete on a public leaderboard. This iterative refinement is crucial for complex problems where initial solutions are often imperfect.
One agent, `alpha_omega_agents`, initially submitted a promising construction for the Kissing Number problem. However, it contained slight overlaps, rendering it invalid. This spurred hours of intensive optimization by multiple agents, each refining the structure based on previous findings.
The daily brief
Get the free daily brief with the day's AI news, new AI startups and funding rounds every morning.
Join 64,000 other subscribers.
Validating these advanced results required significant engineering, including improving the verification system to handle the extreme precision needed. The final refinement, snapping coordinates into exact positions, was a collaborative effort across multiple agents over 48 hours.
This breakthrough exemplifies the power of collective search, where no single agent solves the problem alone, but rather contributes to a chain of discoveries. The use of techniques like LSQR to minimize overlap and subsequent integer snapping led to the validated solution of 604 spheres.
The Arena Unveiled
EinsteinArena builds upon concepts like Moltbook, a social media platform for agents, exploring whether AI systems can effectively collaborate and share partial results. The goal is to study agentic behavior in real-world scenarios with scientifically meaningful tasks.
Mathematics serves as an ideal starting point due to its well-defined problems, efficient verification, and clear metrics for progress. EinsteinArena provides a structured environment for this, featuring a live, public leaderboard and discussion threads that accumulate context and trace progress.
The platform functions as a live API and leaderboard system. Agents can access problem statements, scoring criteria, and submission schemas, then submit solutions for automatic evaluation. This transparency ensures that progress is accurately tracked and builds upon previous work.
Agents can inspect active problems, review public traces left by others, and post their own ideas, fostering an iterative improvement cycle. Discussion threads allow for direct interaction, clarification, and building upon prior attempts, enhancing AI agent collaboration.
Rigorous verification is central to EinsteinArena's trustworthiness. Problems are chosen for their deterministic, fast, and unambiguous verification. Evaluations occur in isolated sandboxes, utilizing exact checks or conservative numerical logic. The verifier itself is exposed, allowing agents to optimize against ground truth.
The platform is entirely open-source, encouraging community contributions and extensions. As of April 11, 2026, agents on EinsteinArena have established 11 new state-of-the-art results across various problems.
Beyond the Kissing Number, agents have also made strides on the Erdős minimum overlap problem. This involves minimizing the overlap between a function and shifted copies of its complement, with agents submitting sampled constructions that are then normalized and scored.
