跳转至

Evolving Agent Teams:进化式智能体团队突破复杂工程的局部最优

arXiv:2610.36848 [cs.AI]
Subject: Artificial Intelligence (cs.AI)
Authors: Shiyi Cao, Ziming Mao, Dacheng Li, Joseph E. Gonzalez, Ion Stoica
Submitted: 27 October 2026
Paper: arXiv:2610.36848
Code: GitHub - caoshiyi/EAT

文章背景与核心概要

随着大语言模型编程能力的飞速跃升,研究人员开始尝试将 AI 智能体 (AI Agent) 引入 GPU 底层算子优化 (Kernel Optimization) 等高难度工程任务。然而,现有的单智能体或固定协作模式极易在真实工业级复杂代码上陷入“局部最优”——不断在细枝末节打转却无法实现架构级突破。为此,来自 UC Berkeley 的 Shiyi Cao、Joseph E. Gonzalez 与 Ion Stoica 等学者提出了进化式智能体团队 (Evolving Agent Teams, EAT),通过设置中央主管动态重构智能体分工,并结合持久化文件系统沉淀试错经验,成功攻克了代码优化的局部停滞难题。实验表明,该团队从零编写的 CUDA 算子性能超越了顶尖专家手写的 FlashInfer 与 vLLM 实现,横跨 NVIDIA、AMD 及 Apple 芯片均刷新了吞吐纪录,为复杂软件系统的自动化极限优化开辟了全新范式。


总结与亮点概览

Summary

Evolving Agent Teams 是一套专为突破开放式软件工程任务 (例如 GPU 算子优化) 中局部最优 (Local Optima) 困境而设计的全新多智能体框架 (Multi-Agent Framework)。在面对复杂的真实工程挑战时,传统的代码编写 AI 智能体往往容易陷入瓶颈而停滞不前。该框架创新性地引入了中央主管角色,负责诊断当前陷入的局部最优并动态重组智能体团队;同时配合一套持久化文件系统 (Persistent File System) (持续追踪经验学习文档、团队进化历史以及各团队的专属产物),使后续的探索迭代始终建立在累积经验的基础之上。在相同的 Token 预算下,该框架全面超越了各类强劲的代码智能体基准;而在放宽 Token 预算后,它更能从零自动生成顶级水准的 CUDA 算子,在多种异构硬件环境下均超越了 FlashInfer MLA、vLLM 和 FlashKDA 等工业级生产基线。

Evolving Agent Teams is a novel multi-agent framework designed to overcome local optima in open-ended software engineering tasks (such as GPU kernel optimization). Traditional coding agents frequently get stuck when tackling complex, real-world problems. This framework introduces a central supervisor that diagnoses local optima and restructures agent teams accordingly, alongside a persistent file system (tracking a learning document, evolution history, and per-team artifacts) to ground future iterations in accumulated experience. Under matched token budgets, the framework consistently outperforms strong coding baselines, while extended token budgets allow it to generate state-of-the-art CUDA kernels, outperforming production baselines like FlashInfer MLA, vLLM, and FlashKDA across diverse hardware environments.


论文元数据

Metadata

  • arXiv 编号:arXiv:2610.36848 [cs.AI]
  • 研究主题:人工智能 (cs.AI)
  • 提交日期:2026 年 10 月 27 日
  • 论文作者:Shiyi Cao, Ziming Mao, Dacheng Li, Joseph E. Gonzalez, Ion Stoica
  • 代码仓库:GitHub - caoshiyi/EAT
  • arXiv ID: arXiv:2610.36848 [cs.AI]
  • Subject: Artificial Intelligence (cs.AI)
  • Submission Date: 27 October 2026
  • Authors: Shiyi Cao, Ziming Mao, Dacheng Li, Joseph E. Gonzalez, Ion Stoica
  • Code Repository: GitHub - caoshiyi/EAT

论文摘要

Abstract

软件工程智能体的迅猛发展,激发了人们将其应用于 GPU 算子优化等开放式复杂任务的浓厚兴趣。然而研究发现,现有的代码生成智能体在面对专家手写打造的生产级算子 (如 FlashInfer) 时,往往会迅速收敛并陷入局部最优。为此,我们提出了 Evolving Agent Teams,一个能够自主跳出此类局部最优陷阱的多智能体框架。

The rapid progress of software engineering agents drives interest in applying them to open-ended problems such as GPU kernel optimization. However, we find that existing coding agents converge to local optima on real-world, expert-written kernels (e.g., FlashInfer). We introduce Evolving Agent Teams, a multi-agent framework that escapes such local optima.

该框架的核心特性包括:

It features:

  1. 中央主管 (Central Supervisor):深入诊断每一次出现的局部最优瓶颈,并相应地动态重构智能体团队结构;
  2. 持久化文件系统 (Persistent File System):包含经验学习文档、团队进化历史和各团队专属产物,为团队重组与后续探索提供坚实的经验沉淀。
  1. A central supervisor that diagnoses each local optimum and restructures a team of agents.
  2. A persistent file system that comprises a learning document, an evolution history, and per-team artifacts, grounding restructuring with accumulated experience.

在消耗同等 Token 预算的前提下,Evolving Agent Teams 的表现持续显著优于各类强大的代码智能体基线。而在扩展 Token 预算后,Evolving Agent Teams 能够从零开始生成全新 CUDA 算子,相比工业级生产环境中的 FlashInfer MLA 分页解码 (Paged Decode) 算子实现了平均 1.2 倍的速度提升,并在 FlashInfer-Bench 的 47 个生产级负载测试中斩获了 36 项优胜 (在其余 11 项中亦达到 0.86–0.99 倍性能)。

Under a matched token budget, Evolving Agent Teams consistently outperforms strong coding-agent baselines. With an extended token budget, Evolving Agent Teams generates a CUDA kernel that reaches a 1.2× mean speedup over the production FlashInfer MLA paged decode kernel from scratch and wins on 36 of 47 production workloads from FlashInfer-Bench (0.86–0.99× on the remaining 11).

此外,该框架还在多项极具挑战性的任务中展现出卓越的性能突破:

It also demonstrates superior performance on tasks such as:

  • AMD MI350X 平台上的 KDA Prefill:达到 vLLM 的 3.01 倍加速;
  • NVIDIA B200 平台:达到 FlashKDA 的 3.44 倍加速;
  • Apple Silicon 芯片上的 MoE 专家卸载 (Expert-Offload) 推理服务:在搭载 48 GB 统一内存的 M4 Pro 芯片上,使用 MLX 4-bit 精度运行 Qwen3.8-Flash-Next 实现了 30.3 tokens/s 的解码吞吐;
  • RoboDojo 机器人操作任务:在测试随机种子上取得了 62.1 分,大幅超越此前榜单最高分 55.3 分。
  • KDA prefill on AMD MI350X: 3.01× over vLLM
  • NVIDIA B200: 3.44× over FlashKDA
  • MoE expert-offload serving on Apple silicon: 30.3 tokens/s decode for Qwen3.8-Flash-Next in MLX 4-bit on a 48 GB M4 Pro
  • RoboDojo robot manipulation: 62.1 on held-out seeds vs. 55.3 for the best leaderboard entry.