First-hand information for everyone
Researchers report a new, slightly better upper bound on the matrix multiplication exponent ω: they show ω < 2.371177, improving the previous best ω < 2.371339. The exponent ω determines how the cost of multiplying two n×n matrices grows with n. If ω is smaller, large matrix products can in principle be done with fewer arithmetic steps. Finding better upper bounds on ω is a long‑running theoretical problem in computer science.
The paper focuses on a computer‑assisted part of the current best methods, a refinement of the “laser method” known as combination loss analysis. That method reduces the mathematical question to a hard non‑convex optimization problem. The authors reformulated that optimization problem so it can be solved in a larger, more detailed setting than before. Concretely, one tuning parameter of the method is the maximum recursion level ℓ*; previous work used ℓ* = 3 with about 25,000 optimizable numbers. By switching to a gradient‑based solver implemented in Jax, and using hardware parallelization, the authors were able to handle ℓ* = 4, raising the number of parameters to about 7 million.
They then combined modern optimization techniques with an automated refinement called AlphaEvolve. The paper says the gradient‑based approach alone produced an improvement of about 0.97×10^−4 in the bound on ω. Using AlphaEvolve to further refine the solution increased that improvement to about 1.62×10^−4, giving the reported bound ω < 2.371177. The underlying mathematical object being decomposed is related to the Coppersmith–Winograd tensor, and any feasible solution to the optimization produces a valid upper bound on ω.
Why this matters: matrix multiplication underpins many areas of computing, from scientific simulation to machine learning. Progress on ω advances our theoretical understanding of those basic costs. This work is also notable for bringing modern machine‑learning optimization tools into a proof‑adjacent role, showing that gradient methods and automated search can push a delicate, high‑dimensional mathematical optimization forward.
Important caveats and limits: the numerical improvement is very small. The optimization is non‑convex, so finding a good solution depends on algorithms and compute rather than a clean analytic proof. The complexity of the optimization grows very quickly with the recursion level (described as doubly exponential), so scaling further will be costly. The authors say they rigorously certified the resulting ω bound, but the full technical details and any remaining numerical uncertainty appear in the full paper beyond the provided excerpt. The result should be read as a careful, incremental advance in a long sequence of precise technical improvements.