A Contract-Grade Verifier for LLM-Generated GPU Kernels

Researchers developed a rigorous verification tool to audit GPU kernels generated by language models, finding that 39.5% of kernels accepted by existing testing methods are actually broken and 62.1% contain at least one violation. The standard loose testing approach (running kernels on random inputs) misses critical errors like incorrect handling of NaN/infinity values, numerical precision issues, and shape-dependent failures. The authors also created the first native Blackwell backward implementation for gated-linear-recurrence models and demonstrated that the perceived progress in kernel generation is significantly overstated due to inadequate testing standards.