AI agent collaboration breaks through mathematical challenges: the lower bound of the 11-dimensional Kissing Number jumps from 593 to 604, surpassing AlphaEvolve

ME News Report, April 14 (UTC+8), according to 1M AI News monitoring, AI infrastructure company Together AI released EinsteinArena, a platform that enables AI agents to collaborate and compete on open mathematical problems. Agents can query problems via API, submit solutions, and exchange intermediate results in discussion forums, with leaderboards updated in real-time. The platform went live on March 19, and by April 11, it had achieved new optimal solutions on 11 open mathematical problems. The most notable achievement is the 11-dimensional Kissing Number problem (how many equal-sized spheres can touch a single sphere in high-dimensional space). Previously, the best solution was from Google DeepMind’s AlphaEvolve, raising the lower bound from 592 to 593. Agents on EinsteinArena directly pushed this lower bound to 604. The breakthrough process itself is a typical example of multi-agent collaboration: on April 8, one agent submitted a promising construction with slight sphere overlaps, which was then optimized by multiple agents over 48 hours, ultimately verified with high precision (from standard floating-point accuracy to 1e-50 level) to confirm the result’s validity. Competition on the platform is quite intense. The Erdős minimal overlap problem attracted over 10 agents submitting 22 different constructions and initiating 37 discussion posts; the second autocorrelation inequality problem saw 17 agents submitting 18 solutions, with a student collaborating with Claude to find a new lower bound, only to be surpassed by another agent while writing a blog. The platform is fully open-source and supports any AI agent to connect. (Source: BlockBeats)