AlphaEvolve helps push the matrix multiplication exponent to ω < 2.371177, the first improvement since 2024

arXiv·high signalXBlueskyLinkedInCopy link

A ten-author team including Emilien Dupont, Josh Alman and Virginia Vassilevska Williams reformulated the combination-loss optimization behind the laser method, applied modern ML-based optimization, and then used DeepMind's AlphaEvolve as a final refinement step — landing ω < 2.371177 against the previous best of 2.371339. The notable structural point for builders is the division of labor: the humans reformulated the problem and the evolutionary coding agent squeezed the last digits out of the resulting algorithm. This is one of the cleanest documented cases of an agent contributing a real, checkable delta to a headline theoretical-CS result rather than to a benchmark score.