Evolving Agent Teams wrote a CUDA kernel hitting 1.2x the speed of production FlashInfer and won 36 of 47 FlashInfer-Bench workloads
AspireAI community · arXiv (Evolving Agent Teams authors)
*Coding agents improve on ordinary software tasks but converge to local optima on real-world expert-written kernels such as FlashInfer.* Evolving Agent Teams is a multi-agent framework built to escape that: a central supervisor diagnoses each local optimum and restructures a team of agents, while a persistent file system - a learning document, an evolution history and per-team artifacts - grounds each restructuring in accumulated experience. Under a matched token budget it consistently beats strong coding-agent baselines.
Given an extended token budget the team generates a CUDA kernel reaching 1.2x mean speedup over the production FlashInfer MLA paged decode kernel from scratch, and it wins on 36 of 47 production workloads from FlashInfer-Bench, sitting at 0.86-0.99x on the remaining 11. The same approach reports 3.01x over vLLM for KDA prefill on AMD MI350X and 3.44x over FlashKDA on NVIDIA B200, plus MoE expert-offload serving on Apple silicon at 30.3 tokens per second decode for Qwen3.8-Flash-Next in MLX 4-bit on a 48 GB M4 Pro, and 62.1 on held-out RoboDojo manipulation seeds against 55.3 for the best leaderboard entry. The code repository linked from the abstract was still returning HTTP 404 when checked, so the paper remains the primary source.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.