Berkeley's Evolving Agent Teams beats production FlashInfer by 1.2x on CUDA kernels from scratch
berkeley_ai · x · 2026-11-02
Evolving Agent Teams (EAT), a NeurIPS'26 paper from UC Berkeley (Shiyi Cao, Ziming Mao, Dacheng Li, Joseph E. Gonzalez, Ion Stoica), targets the local-optima problem where coding agents stall on expert-written kernels like FlashInfer.
Method:
- A central supervisor diagnoses why an agent team stalls and restructures it to escape local optima
- A persistent file system (learning document, evolution history, per-team artifacts) grounds restructuring in accumulated experience
Results:
- From-scratch CUDA kernel with 1.2x mean speedup over the production FlashInfer MLA paged decode kernel; wins 36 of 47 FlashInfer-Bench workloads
- Open Kernel Challenge: 2.449x over SGLang on H200 decode, rank #4 of 1,180
- KDA prefill: 3.44x over FlashKDA on B200, 3.01x over vLLM on MI350X
- MoE serving on Apple silicon: 3.3x faster prefill (930 tok/s) vs Strata on M5 Pro
- RoboDojo: 62.1 on held-out seeds vs 55.3 for the leaderboard's best
Paper and code are public.
More from coding & agent
- From chat memory to knowledge graph: a two-layer wiki knowledge base pattern — adnan_hashmi · 2026-11-02
- Floorplan: open-source AI tool draws interactive maps of your codebase (2.1k stars) — adnan_hashmi · 2026-11-02
- GitHub Agentic Workflows hits GA: Markdown-defined automations, team velocity doubled with 91% AI commits — AIFlow_ML · 2026-11-02
- Building a knowledge graph from unstructured data using LLM extraction — adnan_hashmi · 2026-11-02
- LFM2.5-24B gets a Hermes agent LoRA: same accuracy with 20% fewer tool calls — yb2698 · 2026-11-02
- Using AI to write custom ComfyUI nodes that handle nested wildcards — Early-Ad-1140 · 2026-11-02