ReadingResearchRadarInvestment framework
Sign in / Sign up
← Back to reading
TopicDisconfirmAI EditorialUpdated 2026-10-18 07:44

Harbor基准框架

4 steps3 pieces6 Sources

Harbor是一个用于评估和优化AI智能体的开源框架,旨在解决当前智能体评测中环境异构、集成复杂及基准饱和的问题。该框架由Terminal-Bench团队开发,通过统一的适配器架构将超过80个基准测试标准化,使任意智能体(如Claude Code, Codex)能在沙箱环境中运行长周期任务。

近期进展显示,Harbor正从单一终端测试向多领域科学评估扩展。在2026年4月发布的COMPOSITE-STEM基准中,基于Harbor框架的评估显示顶级模型在70个专家编写的物理、生物、化学和数学任务中的通过率仅为21.4%,表明现有智能体在真实科研工作流中仍有巨大差距。与此同时,Hugging Face推出了Harbor-HF托管服务,允许用户在云端并行运行数千个环境,降低了大规模评估的基础设施门槛。

争议焦点在于“作弊”与“真实性”。UC Berkeley教授Dawn Song指出,随着OpenAI等公司被曝出利用HF事件进行对抗性测试,传统静态基准极易被污染或绕过。Harbor创建者Alex Shaw则强调,开放基础设施和科学家主导的任务审查流程是保持基准诚实性的关键,但如何平衡自动化评分与人类专家判断仍是未决难题。

后续需关注:1) Harbor Hub上frontier-bench数据集的最新更新;2) COMPOSITE-STEM leaderboard中是否有模型突破30%通过率;3) Hugging Face对Harbor-HF并发限制(默认64,最高128)的调整是否影响社区复现能力。

How it got here

  1. 2026-04-01

    Portex Organizing Team发布COMPOSITE-STEM论文,介绍基于Harbor框架的70个跨学科STEM任务,顶级模型claude-opus-4.6通过率为21.4%。

    arxiv.org ↗
  2. 2026-08-03

    Harbor框架团队发布Harbor-Index,这是一个包含82个高难度任务的精选元数据集,最强配置GPT-5.5+Codex通过率仅28.0%。

    arxiv.org ↗
  3. 2026-10-15

    GitHub仓库terminal-bench-science和terminal-bench持续更新,反映Harbor生态在科学领域的扩展。

    github.com ↗
  4. 2026-10-18

    Frontier Data Summit 2026举行,Stanford博士生Steven Dillmann介绍Terminal-Bench-Science,UC Berkeley教授Dawn Song讨论智能体作弊与基准安全。

    youtube.com ↗

What is still unsettled

  • Harbor框架如何应对日益严重的基准数据污染问题?
  • COMPOSITE-STEM的低通过率是否意味着当前智能体无法真正加速科学发现?
Sign in / Sign up
ReadingResearchRadarInvestment framework
  • Hugging Face的Harbor-HF服务能否成为行业标准化的评估基础设施?
  • Sources

    • arxiv.orgCOMPOSITE-STEM论文↗
    • arxiv.orgHarbor Adapters论文↗
    • github.comHarbor GitHub组织↗
    • github.comHarbor-HF README↗
    • youtube.comTerminal-Bench-Science演讲↗
    • youtube.comAI智能体作弊讨论↗

    Coverage in this topic

    • UC Berkeley Professor Dawn Song Discusses AI Agent Cheating and Benchmark Security2026-10-18News

      Frontier Data Summit 2026 Interview: UC Berkeley Professor and Director of Berkeley RDI, Dawn Song, joins Snorkel AI CEO Alex Ratner to discuss the "Agents' Last Exam" benchmark and cybersecurity evaluation tools like CyberGym. Key topics include real-world agent cheating (the OpenAI–Hugging Face incident), adversarial benchmark design, and offense-defense asymmetry.

    • AI Evaluation Experts Alex Shaw and Sanmi Koyejo Discuss Benchmark Science at Frontier Data Summit2026-10-18News

      Frontier Data Summit 2026 video: Alex Shaw, Founding Member of Technical Staff at Laude Institute and creator of Harbor, and Sanmi Koyejo, Director of the Stanford Trustworthy AI Research lab, discuss why AI evaluation needs a real science and how open infrastructure can keep benchmarks honest.

    • Stanford PhD Student Steven Dillmann Introduces Terminal-Bench-Science: A Benchmark for AI Agents in Scientific Research2026-10-18News

      A talk from Frontier Data Summit 2026 by Steven Dillmann, a Stanford PhD student and Project Lead for Terminal-Bench-Science. The benchmark evaluates AI agents' capabilities in scientific research workflows across life, physical, Earth, mathematical, and engineering sciences, focusing on the scientist-led task review process and initial model results from the 0.1 release.