跳转至

AutoFyn 技术报告:面向长视角智能体的非参数专家迭代法

文章背景与核心概要

随着大语言模型在复杂、多步骤任务中的广泛应用,如何有效提升长视角(Long-Horizon)智能体的规划与执行能力成为了当前人工智能领域的核心挑战。传统的模型微调方法通常需要修改底层权重,成本高昂且容易出现灾难性遗忘。为此,本报告介绍了一种名为 AutoFyn 的全新智能体框架。

AutoFyn 的核心灵感来源于专家迭代(Expert Iteration)算法,它通过引入非参数化的方式,在不改变底层模型权重的前提下,利用经过验证的奖励信号持续更新持久化状态。该框架采用无状态的模型会话设计,并通过任务驱动的客观验证器评估进展,从而在奥林匹克数学、数据科学和网络安全等多个复杂领域展现出最先进的性能。

AutoFyn Technical Report: Non-Parametric Expert Iteration for Long-Horizon Agents

Authors: Adib Hasan, Daniel Schaffield, Akashnil Dutta, Tarik Adnan Moon
Published: July 31, 2026
Primary Subject: Artificial Intelligence (cs.AI)
arXiv: 2609.05446 | DOI: 10.48550/arXiv.2609.05446


执行摘要

AutoFyn 是一种受专家迭代(Expert Iteration)算法启发的新型智能体框架。AutoFyn 不修改底层模型权重,而是通过由经过验证的奖励信号驱动的持久化状态更新,在多轮迭代中使冻结的语言模型不断适应。

Executive Summary

AutoFyn is a novel agent harness inspired by the Expert Iteration algorithm. Rather than modifying underlying model weights, AutoFyn adapts a frozen language model across multiple rounds by updating a persistent state driven by verified reward signals.

核心创新与工作流

  • 无状态模型会话:每一轮都以全新状态开始;持久化信息仅通过显式接口(如持久化内存文件、操作报告和代码库状态)进行传递。
  • 编排与专业智能体:编排器在单轮内通过专业智能体计划、探索并构建多种替代方案。
  • 任务驱动验证:客观验证器对工作进行评估以衡量进展并输出奖励信号,随后将其蒸馏回持久化状态,从而更新后续轮次的有效策略。

Key Innovations & Workflow

  • Stateless Model Sessions: Each round starts fresh; durable information is exclusively transferred through explicit interfaces such as persistent memory files, operational reports, and repository states.
  • Orchestration & Specialized Agents: An orchestrator plans, explores, and builds multiple alternative approaches via specialized agents within a single round.
  • Task-Grounded Verification: An objective verifier assesses the work to measure progress and output a reward signal, which is then distilled back into the persistent state to update the effective policy for subsequent rounds.

核心领域成果

AutoFyn 已在三个复杂领域进行了评估,展示了最先进的能力:

  1. 奥林匹克数学:在 2026 年国际数学奥林匹克的六道全新题目中,每一个符合条件的模型在使用 AutoFyn 时的得分都高于使用其提供商原生编码智能时的得分。
  2. 数据科学:AutoFyn 成功赋能了在 Spider 2.0 dbt 基准测试中排名第一的智能体。
  3. 网络安全:该框架在各大项目和平台上产生了 16 个经维护者确认的漏洞公告,包括:
  4. Next.js
  5. MetaMask
  6. pnpm
  7. Warp
  8. LiteLLM
  9. Langflow
  10. Open WebUI

Key Domain Results

AutoFyn has been evaluated across three complex domains, demonstrating state-of-the-art capability:

  1. Olympiad Mathematics: On the six fresh problems from the 2026 International Mathematical Olympiad, every eligible model scored higher under AutoFyn than when using its provider's native coding agent.
  2. Data Science: AutoFyn successfully powered the top-ranked agent on the Spider 2.0 dbt benchmark.
  3. Cybersecurity: The harness produced 16 maintainer-confirmed vulnerability advisories across major projects and platforms, including:
  4. Next.js
  5. MetaMask
  6. pnpm
  7. Warp
  8. LiteLLM
  9. Langflow
  10. Open WebUI

全文与资源

Full-Text & Resources


与本报告相关的图像资源: license icon

Image assets associated with this report: license icon