Gimlet Labs describes formal verification research for AI-generated kernels
Gimlet Labs →Gimlet Labs presented an early research prototype for checking whether optimized AI kernels preserve the behavior of reference implementations. Its tensor algebra equivalence checker complements numerical testing by comparing symbolic representations of PyTorch models and optimized code, including Triton kernels. The company described examples in which conventional tests passed but formal verification exposed differences in functionality. Its process translates computations into a common representation and uses the Z3 solver to seek a counterexample or establish equivalence within the supported formulation. Gimlet explicitly characterized the work as an early prototype. The research addresses correctness challenges created by generating and optimizing kernels for heterogeneous inference hardware.