Direct discovery · Computer science · Mathematics
AlphaTensor uses reinforcement learning to discover matrix multiplication algorithms
Provably correct matrix-multiplication algorithms found through reinforcement-learning search.
Summary
The Nature paper reports AlphaTensor, a deep reinforcement learning system that frames matrix multiplication algorithm discovery as a single-player game. It found provably correct algorithms with improved complexity for several matrix sizes and explored variants optimized for hardware runtime.
AI role
Framed algorithm discovery as a game and searched tensor decompositions for improved matrix multiplication methods.
Narrative role
AlphaTensor is a major algorithm-discovery event because AI produced checkable computational results in a foundational area.
Caveat
The result concerns specific matrix sizes and runtimes; it does not broadly automate algorithm design.