AI Judge Spots FlashInfer Edge Cases in Attention

Josh Tobin, co-founder of Recursive Superintelligence, posted that his group's automated research tool found edge cases in FlashInfer that could affect inference performance in vLLM and SGLang. The team had previously built a reward hacking judge for performance optimization tasks. Tobin stated the judge identified the problems while running on fresh autoresearch experiments and that the group helped produce a fix. The post links to an arXiv paper on Flash Attention stability and a FlashInfer GitHub pull request addressing extreme negative logits in masked softmax.