Evolutionary coding agent that found a 4×4 complex matrix multiplication procedure using 48 scalar multiplications, the first gain over Strassen in 56 years.
An evolutionary coding agent that orchestrates a pipeline of LLMs to improve an algorithm by making direct changes to the code, guided by feedback from automated evaluators.
Its search found a procedure multiplying two 4×4 complex-valued matrices with 48 scalar multiplications — the first improvement over Strassen's algorithm in that setting after 56 years.
Applied to Google's computational stacks, it improved data-center scheduling, found a simplification in hardware accelerator circuit design, and accelerated training of the LLM underpinning itself.