arXivDaily arXiv每日学术速递 周一至周五更新
​通知:联邦节假日调休:哥伦布日 / 原住民日(Columbus Day / Indigenous Peoples' Day),今日(13日)无更新。
arXiv 2504.01848cs.AIcs.CL

PaperBench:评估人工智能复制人工智能研究能力的基准

PaperBench: Evaluating AI's Ability to Replicate AI Research

OpenAI(开放人工智能公司)
Giulio Starace, Oliver Jaffe, Dane Sherburn, James Aung, Jun Shern Chan, Leon Maksin, Rachel Dias, Evan Mays, Benjamin Kinsella, Wyatt Thompson, Johannes Heidec… 展开作者

Giulio Starace, Oliver Jaffe, Dane Sherburn, James Aung, Jun Shern Chan, Leon Maksin, Rachel Dias, Evan Mays, Benjamin Kinsella, Wyatt Thompson, Johannes Heidecke, Amelia Glaese, Tejal Patwardhan

AI总结:

本文提出 PaperBench 基准,通过 20 篇 ICML 2024 论文的 8,316 个可评分子任务评估 AI 智能体复制前沿研究的能力,并开发 LLM 评判器,结果显示最佳模型复制得分仅 21.0%,尚未超越人类基线。

摘要:

我们介绍了 PaperBench,这是一个用于评估 AI 智能体复制最先进 AI 研究能力的基准。智能体必须从零开始复制 20 篇 ICML 2024 Spotlight 和 Oral 论文,包括理解论文贡献、开发代码库以及成功执行实验。为了实现客观评估,我们开发了评分细则,将每个复制任务按层次分解为更小的子任务,并给出明确的评分标准。PaperBench 总共包含 8,316 个可单独评分的任务。评分细则与每篇 ICML 论文的作者共同制定,以确保准确性和现实性。为了实现可扩展的评估,我们还开发了一个基于 LLM 的评判器,根据评分细则自动对复制尝试进行评分,并通过创建一个单独的评判器基准来评估我们评判器的性能。我们在 PaperBench 上评估了多个前沿模型,发现表现最好的测试智能体——使用开源脚手架(scaffolding)的 Claude 3.5 Sonnet (New)——达到了 21.0% 的平均复制得分。最后,我们招募了顶尖的机器学习博士生来尝试 PaperBench 的一个子集,发现模型尚未超越人类基线。我们开源了我们的代码(https://github.com/openai/preparedness),以促进未来在理解 AI 智能体的 AI 工程能力方面的研究。

英文摘要:

We introduce PaperBench, a benchmark evaluating the ability of AI agents to replicate state-of-the-art AI research. Agents must replicate 20 ICML 2024 Spotlight and Oral papers from scratch, including understanding paper contributions, developing a codebase, and successfully executing experiments. For objective evaluation, we develop rubrics that hierarchically decompose each replication task into smaller sub-tasks with clear grading criteria. In total, PaperBench contains 8,316 individually gradable tasks. Rubrics are co-developed with the author(s) of each ICML paper for accuracy and realism. To enable scalable evaluation, we also develop an LLM-based judge to automatically grade replication attempts against rubrics, and assess our judge's performance by creating a separate benchmark for judges. We evaluate several frontier models on PaperBench, finding that the best-performing tested agent, Claude 3.5 Sonnet (New) with open-source scaffolding, achieves an average replication score of 21.0%. Finally, we recruit top ML PhDs to attempt a subset of PaperBench, finding that models do not yet outperform the human baseline. We open-source our code (https://github.com/openai/preparedness) to facilitate future research in understanding the AI engineering capabilities of AI agents.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑