Reinforcement Learning Breaks 50-Year Matrix Multiplication Record

Artificial intelligence has uncovered novel algorithms for matrix multiplication, surpassing human mathematical benchmarks established over half a century ago.

Researchers have demonstrated that reinforcement learning models can discover faster procedures for fundamental arithmetic operations. The milestone targets matrix multiplication, a computational bedrock supporting 3D graphics, physics simulations, and the training of neural networks. For decades, efficiency gains in this domain relied almost exclusively on human mathematical intuition.

The historical baseline dates to 1969, when mathematician Volker Strassen proved multiplying two two-by-two matrices required seven multiplication steps rather than eight. That discovery overturned centuries of standard algebraic practice. Subsequent progress on larger matrices proved exceedingly difficult. For four-by-four matrices, mathematicians remained stalled for more than fifty years at a threshold of forty-nine operations.

The newly unveiled system approached the challenge by converting the algebra of tensor decomposition into a single-player game. In this formulation, the state of the board represents the computational problem to be solved. Each legal move corresponds to an allowable algebraic operation. By evaluating millions of potential paths through self-play, the system identified step combinations that human researchers had overlooked.

The resulting algorithm executes four-by-four matrix multiplication in forty-seven steps. Although saving two steps appears modest in isolation, these routines run billions of times per second across global computing infrastructure. Even fractional reductions in computational complexity yield compounding efficiency gains across commercial hardware.

Beyond four-by-four grids, the system discovered thousands of functioning algorithms for various matrix dimensions. Certain configurations proved tailored to specific hardware architectures, demonstrating that automated search can optimise code for real-world silicon alongside pure mathematical theory.

The breakthrough reframes how computer scientists approach algorithmic design. Rather than relying entirely on manual derivation, laboratories can deploy targeted search agents to navigate complex mathematical spaces.

If learning algorithms can systematically outmanoeuvre human theorists on foundational arithmetic, which long-standing mathematical bottlenecks will yield next?