Researchers improved the theoretical upper bound on the matrix multiplication exponent to 2.371177, down from the prior best of 2.371339, by reformulating the optimization problem behind combination-loss analysis and solving it with gradient-based optimization in JAX. The new approach handles a much larger recursion level using roughly 7 million parameters, versus about 25,000 previously, aided by the Sinkhorn-Knopp algorithm for entropy-maximizing distributions and automatic differentiation for numerical stability. Google DeepMind’s AlphaEvolve system was also used to evolve the optimization algorithm itself, contributing a further incremental improvement.
