Who&When Pro: Can LLMs Really Attribute Failures in AI Agents?
Proposes Who&When Pro, using error injection to generate 12,326 high-quality failure traces across multimodal tasks for automated failure attribution.
Key Findings
Methodology
This work employs a controlled error injection pipeline, where successful agent trajectories from 15 frameworks across 26 benchmarks are selectively corrupted at specific steps using a conditioned generative model. The process involves selecting a decisive step based on the Decisive Error framework, generating an error action via a context-aware prompt, and performing warm-start replays to ensure environment and state consistency. This approach avoids drift issues common in re-rollout methods, producing high-fidelity failure labels. The pipeline includes manual validation, resulting in 12,326 labeled failure trajectories covering text, image, and video modalities, with detailed failure taxonomy and multi-stage filtering to ensure label accuracy.
Key Results
- The benchmark includes 12,326 failure trajectories, spanning 26 benchmarks, 9 task categories, and 3 modalities, significantly expanding the scope of failure attribution datasets. Models like GPT-5.4 achieve 73.9% step localization accuracy on text, but performance drops to 50% on video, highlighting modality-specific challenges. Error mode classification F1 scores hover around 22%, indicating room for improvement. Cross-modal analysis shows that visual cues improve failure mode detection, while longer sequences (>12K tokens) reduce step localization accuracy to 50%.
- Evaluation of multiple models reveals that full-trajectory attribution outperforms incremental methods, with GPT-5.4 leading in step localization, while Qwen3.5-122B excels in agent identification. The results demonstrate that failure attribution remains a complex task, especially in multimodal and long-sequence scenarios, with accuracy significantly decreasing as trace length increases.
- Experimental results underscore the difficulty of fine-grained failure diagnosis, especially in video and lengthy traces, where models struggle with causal tracing. The study highlights the importance of multimodal cues and comprehensive trajectory analysis for improving failure attribution performance.
Significance
This benchmark advances the field by providing a large-scale, multimodal dataset for automated failure attribution, addressing limitations of previous text-only datasets. It enables systematic evaluation of models’ ability to identify failure sources across diverse scenarios, fostering progress toward more reliable and self-aware AI agents. The methodology supports the development of autonomous systems capable of self-diagnosis and correction, crucial for deploying AI in safety-critical applications like autonomous driving, robotics, and healthcare. By bridging the gap between model capability and failure detection, this work paves the way for more robust, trustworthy AI systems that can learn from their mistakes without human intervention.
Technical Contribution
The core technical innovation lies in integrating a decisive-error framework with a context-aware, error-conditioned generation pipeline that prevents drift issues common in re-rollout approaches. The pipeline’s ability to produce high-fidelity, step-level failure labels across multiple modalities is a significant advancement. Additionally, the construction of a comprehensive failure taxonomy and the multi-stage filtering process enhance label quality. The approach also introduces a systematic evaluation protocol comparing full-trajectory, incremental, and binary search attribution methods, demonstrating the superiority of the former in complex, multimodal environments. These contributions collectively enable scalable, accurate failure attribution that supports model self-evolution and debugging.
Novelty
This work is the first to establish a large-scale, multimodal failure attribution benchmark based on automated, error-injected trajectories, moving beyond prior text-only datasets. The innovative use of a context-aware, warm-start re-execution process to generate high-fidelity labels distinguishes it from existing re-rollout or simulation-based methods. The comprehensive taxonomy and multi-task setup further enrich the failure understanding, making it a pioneering effort in multi-modal failure analysis. Compared to prior work, it offers a scalable, high-quality dataset with detailed failure annotations across diverse scenarios, facilitating future research in failure diagnosis and model robustness.
Limitations
- The method relies on predefined failure types and error models, which may not cover all real-world failure modes, limiting generalization.
- Performance drops significantly on long sequences (>12K tokens) and complex video tasks, indicating the need for better long-context modeling and multimodal fusion techniques.
- Error injection, while avoiding drift, may not fully replicate real-world failures, potentially limiting the ecological validity of the labels.
Future Work
Future directions include developing adaptive error models that better mimic real failures, enhancing long-sequence and multimodal processing capabilities, and integrating self-reflection mechanisms for autonomous failure diagnosis. Expanding the failure taxonomy with unsupervised discovery methods and applying the benchmark to real-world deployment scenarios will further improve robustness. Additionally, exploring reinforcement learning-based self-evolution strategies driven by these high-quality labels could lead to more autonomous, self-improving agents.
AI Executive Summary
The rapid growth of multimodal AI agents has heightened the need for precise failure detection and attribution. Traditional approaches rely heavily on manual annotations, which are labor-intensive and limited in scope, especially across diverse modalities like images and videos. Addressing this gap, the present work introduces Who&When Pro, a large-scale benchmark built upon a novel error injection pipeline. This pipeline systematically corrupts successful agent trajectories at carefully selected steps, generating failure cases with high-fidelity, gold-standard labels. The core innovation lies in the use of a context-aware, warm-start re-execution process that prevents drift, ensuring the accuracy of failure attribution across text, image, and video modalities.
By applying this approach, the authors assembled 12,326 failure trajectories covering 26 benchmarks and multiple agent frameworks. These trajectories encompass a broad spectrum of failure modes, including perception, reasoning, planning, and coordination errors, providing a comprehensive dataset for evaluating failure attribution models. Experimental results reveal that current models, such as GPT-5.4, achieve up to 73.9% accuracy in step localization on text tasks, but performance diminishes significantly in multimodal and long-sequence scenarios, highlighting ongoing challenges.
The benchmark’s significance extends beyond mere evaluation. It offers a foundation for developing more robust, self-diagnosing AI systems capable of learning from failures without human intervention. The detailed failure taxonomy and multi-stage filtering ensure high label quality, fostering research that bridges the gap between model capability and fault detection. Looking ahead, future work will focus on refining error models, improving long-context understanding, and integrating self-reflective mechanisms to enable autonomous system improvement. Overall, Who&When Pro marks a pivotal step toward trustworthy, self-evolving AI agents capable of operating reliably in complex, real-world environments.
Deep Analysis
Background
近年来，随着大规模预训练模型（如GPT、BERT）在自然语言处理中的突破，智能代理逐步向多模态、多任务方向发展。代表性工作包括ReAct（Yao et al., 2022）、Toolformer（Schick & Schütze, 2023）等，推动了模型在推理、规划、工具调用等方面的能力提升。然而，随着模型能力增强，失败的隐蔽性也在增加，传统人工检测和有限基准难以满足实际需求。现有失败归因多集中于文本任务（如Zhang et al., 2025c），缺乏跨模态、长序列和大规模自动化标注，限制了其推广应用。
Core Problem
自动失败归因旨在识别系统在哪个步骤、由哪个代理引发了失败，以及失败的具体原因。现有方法多依赖人工标注或有限自动化流程，难以应对多模态、多任务环境中的复杂场景。尤其在视频和图像任务中，缺乏高质量、规模化的标注，导致模型难以准确诊断故障。如何在保证标签真实性的同时实现大规模自动化，成为核心难题。
Innovation
本研究的创新点包括：1）提出基于误差注入的自动化标注流程，确保标签的金标准性质；2）引入Decisive Error定义，有效定位导致任务失败的关键步骤；3）采用温启动机制，避免漂移问题，提升标签一致性；4）构建跨模态、多任务的失败轨迹库，丰富故障类型和场景，支持多样化应用。这些创新突破了传统人工标注的瓶颈，为多模态失败归因提供了新思路。
Methodology
- �� 采集源轨迹：从15个代理框架和26个基准任务中收集成功轨迹，涵盖文本、图像、视频模态。
- �� 失败分类：人工分析失败轨迹，建立多模态故障类型体系，包括感知、推理、规划、验证、协调等。
- �� 误差注入：在成功轨迹中选择特定步骤，利用条件生成模型（如GPT-4）生成误差行为，插入轨迹中。
- �� 温启动回放：从注入点开始，重放轨迹，确保环境状态和上下文一致，避免漂移。
- �� 失败验证：筛选出因误差导致的失败轨迹，确保标签的真实性。
- �� 人工验证：由三名专家对部分样本进行确认，确保标注质量。
Experiments
采用26个基准任务、15个代理框架，构建12326条多模态失败轨迹。评估模型包括GPT-5.4、Qwen3.5-122B等，指标涵盖步骤定位准确率、错误分类F1、联合准确率。通过对比不同归因协议（全轨迹、逐步、二分法），验证全轨迹归因的优越性。还进行了长序列和多模态场景的性能分析，揭示模型在复杂环境中的不足。
Results
模型在文本模态下达到了73.9%的步骤定位准确率，但在视频模态中下降至50%，显示多模态和长序列对归因能力的挑战。不同模型在不同任务中的表现差异明显，GPT-5.4在步骤识别方面表现最佳，Qwen3.5-122B在代理识别上优异。多模态融合提供了更丰富的故障线索，但也带来了更高的复杂性。整体来看，模型在细粒度归因方面仍有提升空间。
Applications
该基准可用于训练和评估未来多模态智能系统的故障检测能力，特别适合自动驾驶、机器人、智能制造等场景。通过自动化生成高质量标签，减少人工成本，加速系统的迭代优化。未来，结合强化学习和自我反思机制，有望实现自主故障诊断和修正，提升系统的安全性和可靠性。
Limitations & Outlook
目前方法依赖于预定义的故障类型和误差模型，难以覆盖所有实际场景中的故障类型。长序列和复杂视频模态下，归因准确率仍有明显下降，表明模型在处理大规模、多模态信息时仍面临挑战。此外，误差注入的模拟可能无法完全反映真实故障场景，限制了标签的真实性和泛化能力。
Plain Language Accessible to non-experts
想象你在一家工厂工作，工厂里有很多不同的机器，每个机器负责不同的任务。有时候，机器会出错，比如误操作或零件损坏。工厂的管理者希望知道是哪台机器出了问题、在哪个环节出错、以及为什么出错。传统的方法是工人逐个检查机器，但费时费力。现在，有一种智能助手可以模拟故障，故意让某台机器出错，然后观察它的反应，帮助找到问题的根源。这就像在工厂里故意让某个环节出错，看看哪个环节最容易出错，帮助工厂更快修复。这个研究就是用类似的方法，让AI系统自己模拟出错，然后自动找到出错的原因和位置，从而提升系统的可靠性。
ELI14 Explained like you're 14
想象你在学校里，有很多不同的老师和学生。有时候，考试会出错，原因可能是题目不清楚、学生没理解，或者老师讲错了。老师们想知道，哪个老师出了错、在哪个题目上出了问题、为什么会出错。以前，老师们要自己检查每个题目，花费很多时间。现在，有个聪明的机器人，可以模拟出错的情况，告诉老师们哪里出了问题。它会在题目中故意设置一些错误，然后观察学生的反应，帮助老师找到真正的问题所在。这个研究也是这样：用电脑模拟出错，然后自动找出错的原因和位置，帮助让系统变得更聪明、更可靠。
Abstract
Automated failure attribution uses LLMs to identify where and why agentic systems fail. As agents become more capable, their failures become subtler, making automated attribution increasingly important. We introduce Who&When Pro, a large-scale benchmark for automated failure attribution in agentic systems. Using a strictly controlled pipeline that injects a failure only after exactly replaying a successful prefix, we construct 12,326 failed trajectories with golden labels across 3 modalities and 26 benchmarks covering various scenarios. Beyond benchmarking, we conduct extensive experiments and analyses, revealing systematic patterns in how models attribute failures across modalities, protocols, and model families, and providing empirical guidance for future automated failure attribution systems.