Tag
ARBOUR 提出一种推理失败归因框架，将推理轨迹表示为推理状态树，以全局视角定位长思维链中的过度思考、思考不足和无依据推断等失败模式。研究发现该树状状态能区分真正导致失败的步骤与风险步骤。
ReCast 提出一种面向归因的步表示学习方法，将冻结 LLM 的隐藏状态转换为可更好区分根因步骤的表示，并发布 ReCast-2K 训练数据集，在四个基准上取得最佳 Hit@1 成绩。
This paper introduces EvoGen-Harness, a generator-agnostic framework that evolves multiple external responsibilities around frozen text-to-image models using a Trace method for failure attribution and coordinated search, achieving significant benchmark gains over existing single-dimension adaptation baselines.
The paper introduces Continual Search, an iterative framework to enhance root-cause attribution for AI agent failures by searching for diagnostic evidence in long-horizon execution traces, and evaluates it on benchmarks including MegaRCA-Mix, showing significant performance improvements.
The paper introduces AgenticRAG-FP, an interventional benchmark for causal failure attribution in agentic RAG systems, analyzing how failures propagate across retrieval hops and evaluating diagnosis performance.
The paper introduces self-diagnosing models that attribute model failures under distribution shift, linking uncertainty estimation with failure attribution.
This paper introduces OAT, a lightweight failure attribution tool for LLM-based agentic systems that trains only on successful trajectories and uses neural controlled differential equations to detect error steps, outperforming expensive baselines by orders of magnitude in speed and accuracy.
Presents Oat, a lightweight unsupervised method for identifying error steps in LLM-based agentic failure trajectories using neural controlled differential equations trained only on successful trajectories. It achieves 200-5000x speedup over prompting baselines with significant F1 improvements in in-domain and out-of-distribution settings.
This paper introduces Who&When Pro, a large-scale benchmark for automated failure attribution in AI agentic systems, with 12,326 failure trajectories across 26 benchmarks, and provides insights into how LLMs attribute failures.
StepFinder is a lightweight framework that uses LLMs only in the feature construction phase to encode execution logs into temporal semantic sequences, then applies parameter-efficient temporal and attention modules for failure attribution in multi-agent systems. It reduces inference time by 79% compared to the fastest LLM-based method on the Who&When benchmark.
SkillAdaptor is a training-free step-level skill adaptation framework with explicit failure attribution for LLM agents, improving performance on WebShop, PinchBench, and Claw-Eval.
This survey paper provides a unified review of LLM-based multi-agent systems, focusing on collaboration, failure attribution, and self-evolution through the LIFE framework, identifying open challenges and proposing a cross-stage research agenda.