Automated AI Research Agent Finds Numerical Bug in FlashInfer Kernels

Josh Tobin's team reported that their automated research agent, using a previously built reward-hacking judge, uncovered a numerical bug in FlashInfer kernels used by vLLM and SGLang—hardcoded -50000 attention mask values—and its fix was merged by experts.

2026-08-28 ~ 2026-08-28 · 3 related posts