TRAIL
收藏资源简介:
TRAIL(跟踪推理和代理问题定位)是一个包含148个标注的AI代理执行跟踪的基准数据集,涵盖推理、执行和规划类别的841个错误。该数据集源自现实世界的软件工程和信息检索任务,挑战了最先进的语言模型,最佳模型的准确率仅为11%,突显了复杂代理工作流跟踪调试的难度。
TRAIL (Tracking Reasoning and Agent Problem Localization) is a benchmark dataset consisting of 148 annotated AI agent execution traces, encompassing 841 errors across three categories: reasoning, execution, and planning. Originating from real-world software engineering and information retrieval tasks, this dataset poses considerable challenges to state-of-the-art language models, as the best-performing model achieves only 11% accuracy, thereby highlighting the inherent difficulty of tracking and debugging complex agent workflows.
TRAIL 基准测试数据集概述
数据集简介
- 名称:TRAIL(Trace Reasoning and Agentic Issue Localization)
- 内容:包含148个标注的AI代理执行轨迹,涵盖841个错误
- 错误类别:推理、执行和规划
- 来源:基于真实世界的软件工程和信息检索任务构建
关键特性
- 挑战性:即使最先进的LLM模型也仅达到11%的准确率
- 目的:用于评估复杂代理工作流中的轨迹调试能力
技术规格
- 安装要求:需通过
pip install -r requirements.txt安装依赖包 - 评估脚本:
run_eval.py支持以下参数:model:兼容litellm的模型IDdata_dir:数据目录路径(默认"data/")output_dir:结果输出目录(默认"results/")max_workers:工作线程数split:数据集分片(可选"GAIA"或"SWE Bench")
结果计算
- 使用
calculate_scores.py脚本计算评估结果 - 输出为存储在
results/目录下的.txt文件
引用信息
bibtex @misc{deshpande2025trail, title={TRAIL: Trace Reasoning and Agentic Issue Localization}, author={Darshan Deshpande and Varun Gangal and Hersh Mehta and Jitin Krishnan and Anand Kannappan and Rebecca Qian}, year={2025}, eprint={2505.08638}, archivePrefix={arXiv}, primaryClass={cs.AI}, url={https://arxiv.org/abs/2505.08638}, }




