AlphaTensor

DeepMind's AlphaZero-based agent that discovers provably correct matrix multiplication algorithms, including a 47-multiplication 4×4 method.

Specifications & Metrics

A deep reinforcement learning agent based on AlphaZero that discovers efficient, provably correct algorithms for multiplying arbitrary matrices.
It found an algorithm multiplying 4×4 matrices in Z₂ with 47 multiplications, outperforming Strassen's two-level algorithm with 7² = 49.
For that matrix size this was, to the authors' knowledge, the first improvement since Strassen's discovery 50 years earlier.