This presentation introduces KernelQuest, a benchmark from Turing Frontier Research Lab for Triton GPU kernel optimization, and a verifier that distinguishes genuine speedups from invalid shortcuts or altered program behavior. Findings from frontier agent evaluations, independent audits, and adversarial testing show why reliable verification is essential and how it must evolve as agents become more capable.

ML4Science-Bench from the Turing Frontier Research Lab evaluates whether autonomous agents can complete scientifically valid machine learning workflows across biology, chemistry, mathematics, and physics. Key takeaways include the gap between runnable code and reliable scientific results, the importance of testing methodology and reproducibility, and the need to assess agent success alongside consistency, cost, and speed.