CertiK and researchers at Columbia University introduced TensorLean and used it to formally verify production-grade AI operators. According to Foresight News, the work applies mathematical proofs to optimized AI computing code, offering model developers stronger correctness checks as they improve training speed and reduce inference costs.
The team evaluated 149 compiled GPU kernel instances and formally verified that 83 met their corresponding Lean specifications. The research covered multiple generations of GPU architecture and included DeepSeek’s DeepGEMM and the FlashAttention operator library. Across six libraries, researchers identified 38 correctness issues, including two in DeepGEMM and 23 in FlashAttention.
TensorLean can be used to recheck code after updates, as AI computing code changes to accommodate performance tuning and hardware. The approach may help reduce computational errors, repeated training work and wasted computing resources.