Berkeley's Evolving Agent Teams Achieve SOTA in CUDA Kernel Optimization

UC Berkeley researchers propose Evolving Agent Teams (EAT), a multi-agent framework for coding tasks, which ranked 4th and set SOTA on the Open Kernel Challenge, producing CUDA kernels that outperform production-level FlashInfer by 1.2x.

2026-11-02 ~ 2026-11-03 · 2 related posts

• Berkeley's Evolving Agent Teams beats production FlashInfer by 1.2x on CUDA kernels from scratch — berkeley_ai · 2026-11-02
• EAT Evolves Agent Teams to Hit SOTA on Kernel Optimization, Ranks #4 in Open Kernel Challenge — berkeley_ai · 2026-11-03