Datasets:

RobinChen2001
/

A-Survey-for-LLM-Agent-Trajectory-Analysis

title stringlengths 5 133	category stringclasses 6 values	subcategory stringclasses 14 values	venue stringlengths 3 27	year stringdate 2025-01-01 00:00:00 2026-01-01 00:00:00	note stringlengths 0 164	paper_url stringlengths 0 46	code_url stringlengths 0 60	dataset_url stringclasses 4 values	project_url stringclasses 3 values	doi_url stringclasses 4 values
TRAIL: Trace Reasoning and Agentic Issue Localization	Failure Taxonomy		arXiv	2025		https://arxiv.org/abs/2505.08638	https://github.com/patronus-ai/trail-benchmark
Aegis: Taxonomy and Optimizations for Overcoming Agent-environment Failures in LLM Agents	Failure Taxonomy		arXiv	2025		https://arxiv.org/abs/2508.19504
Exploring Autonomous Agents: A Closer Look at Why They Fail When Completing Tasks	Failure Taxonomy		ASE	2025		https://arxiv.org/abs/2508.13143
Where LLM Agents Fail and How They Can Learn From Failures	Failure Taxonomy		arXiv	2025		https://arxiv.org/abs/2509.25370	https://github.com/ulab-uiuc/AgentDebug
Why Do Multi-Agent LLM Systems Fail?	Failure Taxonomy		ICLR Workshop	2025		https://arxiv.org/abs/2503.13657	https://github.com/multi-agent-systems-failure-taxonomy/MAST
How Do LLMs Fail In Agentic Scenarios? A Qualitative Analysis of Success and Failure Scenarios of Various LLMs in Agentic Simulations	Failure Taxonomy		NeurIPS	2025		https://arxiv.org/abs/2512.07497
AgentRx: Diagnosing AI Agent Failures from Execution Trajectories	Failure Taxonomy		arXiv	2026		https://arxiv.org/abs/2602.02475
Demystifying the Lifecycle of Failures in Platform-Orchestrated Agentic Workflows	Failure Taxonomy		arXiv	2025		https://arxiv.org/abs/2509.23735v2
AgentEval: DAG-Structured Step-Level Evaluation for Agentic Workflows with Error Propagation Tracking	Failure Taxonomy		arXiv	2026		https://arxiv.org/abs/2604.23581
Who is Introducing the Failure? Automatically Attributing Failures of Multi-Agent Systems via Spectrum Analysis (FAMAS)	Failure Attribution Methods for LLM Agents	Pattern Analysis-Based	ESEC/FSE	2026		https://arxiv.org/abs/2509.13782
Traceability and Accountability in Role-Specialized Multi-Agent LLM Pipelines	Failure Attribution Methods for LLM Agents	Pattern Analysis-Based	arXiv	2025		https://arxiv.org/abs/2510.07614		https://sites.google.com/view/mas-gain2025/home
CORRECT: Condensed eRror Recognition via Knowledge Transfer in Multi-Agent Systems	Failure Attribution Methods for LLM Agents	Pattern Analysis-Based	arXiv	2025		https://arxiv.org/abs/2509.24088
Scope Delineation Before Localization (SDBL)	Failure Attribution Methods for LLM Agents	Pattern Analysis-Based	AAAI	2026		https://arxiv.org/abs/2512.15374	https://github.com/JarvisPei/SCOPE
AgentEval: DAG-Structured Step-Level Evaluation for Agentic Workflows with Error Propagation Tracking	Failure Attribution Methods for LLM Agents	Pattern Analysis-Based	arXiv	2026		https://arxiv.org/abs/2604.23581
ProMAS: Proactive Error Forecasting for Multi-Agent Systems Using Markov Transition Dynamics	Failure Attribution Methods for LLM Agents	Pattern Analysis-Based	arXiv	2026		https://arxiv.org/abs/2603.20260
Which Agent Causes Task Failures and When?	Failure Attribution Methods for LLM Agents	LLM Reasoning-Based	ICML	2025		https://arxiv.org/abs/2505.00212	https://github.com/ag2ai/Agents_Failure_Attribution
Where Did It All Go Wrong? A Hierarchical Look into Multi-Agent Error Attribution (ECHO)	Failure Attribution Methods for LLM Agents	LLM Reasoning-Based	NeurIPS	2025		https://arxiv.org/abs/2510.04886
RAFFLES: Reasoning-based Attribution of Faults for LLM Systems	Failure Attribution Methods for LLM Agents	LLM Reasoning-Based	NeurIPS Workshop	2025		https://arxiv.org/abs/2509.06822
Automatic Failure Attribution and Critical Step Prediction based on Causal Inference (CDC-MAS)	Failure Attribution Methods for LLM Agents	LLM Reasoning-Based	arXiv	2025		https://arxiv.org/abs/2509.08682
Abduct, Act, Predict: Scaffolding Causal Inference for Automated Failure Attribution in Multi-Agent Systems	Failure Attribution Methods for LLM Agents	LLM Reasoning-Based	NeurIPS	2025		https://arxiv.org/abs/2509.10401	https://github.com/ResearAI/A2P
From Flat Logs to Causal Graphs: Hierarchical Failure Attribution for LLM-based Multi-Agent Systems	Failure Attribution Methods for LLM Agents	LLM Reasoning-Based	arXiv	2026		https://arxiv.org/abs/2602.23701
AgentRx: Diagnosing AI Agent Failures from Execution Trajectories	Failure Attribution Methods for LLM Agents	LLM Reasoning-Based	arXiv	2026		https://arxiv.org/abs/2602.02475
CodeTracer: Towards Traceable Agent States for Failure Attribution	Failure Attribution Methods for LLM Agents	LLM Reasoning-Based	arXiv	2026		https://arxiv.org/abs/2604.11641
ERRORPROBE: Towards Self-Improving Error Diagnosis in Multi-Agent Systems	Failure Attribution Methods for LLM Agents	LLM Reasoning-Based	arXiv	2026		https://arxiv.org/abs/2604.17658
AgenTracer: Who Is Inducing Failure in the LLM Agentic Systems?	Failure Attribution Methods for LLM Agents	Model Fine-Tuning-Based	ICLR	2026		https://arxiv.org/abs/2509.03312	https://github.com/bingreeky/AgenTracer
GraphTracer: Graph-Guided Failure Tracing in LLM Agents	Failure Attribution Methods for LLM Agents	Model Fine-Tuning-Based	arXiv	2025		https://arxiv.org/abs/2510.10581
Aegis: Automated Error Generation and Attribution for Multi-Agent Systems	Failure Attribution Methods for LLM Agents	Model Fine-Tuning-Based	arXiv	2025		https://arxiv.org/abs/2509.14295
DoVer: Intervention-Driven Auto Debugging for LLM Multi-Agent Systems	Failure Attribution Methods for LLM Agents	Dynamic Runtime-Based	ICLR	2026		https://arxiv.org/abs/2512.06749			https://mbjinx.github.io/DoVer_Web/
AgentDebug (Where LLM Agents Fail and How They can Learn From Failures)	Failure Attribution Methods for LLM Agents	Dynamic Runtime-Based	arXiv	2025		https://arxiv.org/abs/2509.25370	https://github.com/ulab-uiuc/AgentDebug
TraceElephant: Seeing the Whole Elephant for Failure Attribution in LLM-based Multi-Agent Systems	Failure Attribution Methods for LLM Agents	Dynamic Runtime-Based	ACL	2026		https://arxiv.org/abs/2604.22708	https://github.com/TraceElephant/TraceElephant
Demystifying the Lifecycle of Failures in Platform-Orchestrated Agentic Workflows	Failure Attribution Methods for LLM Agents	Dynamic Runtime-Based	arXiv	2025		https://arxiv.org/abs/2509.23735v2
Aegis: Taxonomy and Optimizations for Overcoming Agent-Environment Failures in LLM Agents	Enhancement, Optimization, and Repair	Structural and Workflow Optimization	arXiv	2025		https://arxiv.org/abs/2508.19504
Maestro: Joint Graph & Config Optimization for Reliable AI Agents	Enhancement, Optimization, and Repair	Structural and Workflow Optimization	arXiv	2025		https://arxiv.org/abs/2509.04642
Failure-Driven Workflow Refinement (CE-Graph)	Enhancement, Optimization, and Repair	Structural and Workflow Optimization	arXiv	2025		https://arxiv.org/abs/2510.10035
Instruction-Level Weight Shaping (ILWS)	Enhancement, Optimization, and Repair	Structural and Workflow Optimization	arXiv	2025		https://arxiv.org/abs/2509.00251
SCOPE: Prompt Evolution for Enhancing Agent Effectiveness	Enhancement, Optimization, and Repair	Agent Internal Optimization	arXiv	2025		https://arxiv.org/abs/2512.15374	https://github.com/JarvisPei/SCOPE
AgentDevel: Reframing Self-Evolving LLM Agents as Release Engineering	Enhancement, Optimization, and Repair	Agent Internal Optimization	arXiv	2026		https://arxiv.org/abs/2601.04620
ReCreate: Reasoning and Creating Domain Agents Driven by Experience	Enhancement, Optimization, and Repair	Agent Internal Optimization	arXiv	2026		https://arxiv.org/abs/2601.11100	https://github.com/zz-haooo/ReCreate
SEAlign: Alignment Training for Software Engineering Agent	Enhancement, Optimization, and Repair	Agent Internal Optimization	ICSE	2026
Trajectory-Informed Memory Generation for Self-Improving Agent Systems	Enhancement, Optimization, and Repair	Agent Internal Optimization	arXiv	2026		https://arxiv.org/abs/2603.10600
Trace2Skill: Distill Trajectory-Local Lessons into Transferable Agent Skills	Enhancement, Optimization, and Repair	Agent Internal Optimization	arXiv	2026		https://arxiv.org/abs/2603.25158
Improving the Efficiency of LLM Agent Systems through Trajectory Reduction (AgentDiet)	Enhancement, Optimization, and Repair	Runtime and Supervisory Optimization	arXiv	2025		https://arxiv.org/abs/2509.23586
Stop Wasting Your Tokens: Towards Efficient Runtime Multi-Agent Systems (SUPERVISOR AGENT)	Enhancement, Optimization, and Repair	Runtime and Supervisory Optimization	arXiv	2025		https://arxiv.org/abs/2510.26585
Wink: Recovering from Misbehaviors in Coding Agents	Enhancement, Optimization, and Repair	Runtime and Supervisory Optimization	arXiv	2026		https://arxiv.org/abs/2602.17037
Process-Centric Analysis of Agentic Software Systems	Enhancement, Optimization, and Repair	Runtime and Supervisory Optimization	OOPSLA	2026
AgentSight: System-Level Observability for AI Agents using eBPF	Trajectory Monitoring, Debugging, and Analysis Tools	System-Level Monitoring and Passive Diagnosis	Workshop	2025		https://dl.acm.org/doi/10.1145/3766882.3767169	https://github.com/eunomia-bpf/agentsight			https://dl.acm.org/doi/10.1145/3766882.3767169
Taming Uncertainty via Automation: Observing, Analyzing, and Optimizing Agentic AI Systems	Trajectory Monitoring, Debugging, and Analysis Tools	System-Level Monitoring and Passive Diagnosis	arXiv	2025		https://arxiv.org/abs/2507.11277
AgentDiagnose: An Open Toolkit for Diagnosing LLM Agent Trajectories	Trajectory Monitoring, Debugging, and Analysis Tools	System-Level Monitoring and Passive Diagnosis	EMNLP	2025		https://aclanthology.org/2025.emnlp-demos.15/	https://github.com/oootttyyy/AgentDiagnose
Agent Trajectory Explorer: Visualizing and Providing Feedback on Agent Trajectories	Trajectory Monitoring, Debugging, and Analysis Tools	System-Level Monitoring and Passive Diagnosis	AAAI	2025		https://doi.org/10.1609/aaai.v39i28.35350				https://doi.org/10.1609/aaai.v39i28.35350
Interactive Debugging and Steering of Multi-Agent AI Systems (AGDebugger)	Trajectory Monitoring, Debugging, and Analysis Tools	Interactive Analysis and Active Debugging	CHI	2025		https://doi.org/10.1145/3706598.3713581	https://github.com/microsoft/agdebugger			https://doi.org/10.1145/3706598.3713581
XAgen: An Explainability Tool for Identifying and Correcting Failures in Multi-Agent Workflows	Trajectory Monitoring, Debugging, and Analysis Tools	Interactive Analysis and Active Debugging	CHI	2025		https://arxiv.org/abs/2512.17896
DiLLS: Interactive Diagnosis of LLM-based Multi-agent Systems via Layered Summary of Agent Behaviors	Trajectory Monitoring, Debugging, and Analysis Tools	Interactive Analysis and Active Debugging	CHI	2026		https://arxiv.org/abs/2602.05446
AgentProcessBench: Diagnosing Step-Level Process Quality in Tool-Using Agents	Datasets and Benchmarks for Failure Attribution and Repair	Process and Span-Level Error Localization	arXiv	2026	1,000 tool-augmented agent trajectories with 8,509 human-labeled step annotations for step-level process quality diagnosis.	https://arxiv.org/abs/2603.14465	https://github.com/RUCBM/AgentProcessBench
CodeTracer / CodeTraceBench: Towards Traceable Agent States	Datasets and Benchmarks for Failure Attribution and Repair	Process and Span-Level Error Localization	arXiv; Hugging Face Dataset	2026	4,316 coding-agent trajectories with human-verified stage- and step-level annotations for failure localization.	https://arxiv.org/abs/2604.11641		https://huggingface.co/datasets/NJU-LINK/CodeTraceBench
TELBench / DRIFT: Where Do Deep-Research Agents Go Wrong? Span-Level Error Localization in Agent Trajectories	Datasets and Benchmarks for Failure Attribution and Repair	Process and Span-Level Error Localization	arXiv; Hugging Face Dataset	2026	1,000 expert-verified deep-research trajectories with semantic spans and harmful error-span annotations.	https://arxiv.org/abs/2606.02060	https://github.com/NJU-LINK/DRIFT	https://huggingface.co/datasets/NJU-LINK/TELBench
ContextBench: A Benchmark for Context Retrieval in Coding Agents	Datasets and Benchmarks for Failure Attribution and Repair	Context Retrieval Benchmarks	arXiv	2026	1,136 issue-resolution tasks across 66 repositories and 8 programming languages, with 4,548 files, 23,116 blocks, and 522,115 lines of human-verified gold contexts.	https://arxiv.org/abs/2602.05892	https://github.com/EuniAI/ContextBench
Who&When	Datasets and Benchmarks for Failure Attribution and Repair	Real-World Failure Collection	ICML	2025	127 trajectories	https://arxiv.org/abs/2505.00212	https://github.com/ag2ai/Agents_Failure_Attribution
TRAIL	Datasets and Benchmarks for Failure Attribution and Repair	Real-World Failure Collection	arXiv	2025	148 trajectories	https://arxiv.org/abs/2505.08638	https://github.com/patronus-ai/trail-benchmark
AgentErrorBench	Datasets and Benchmarks for Failure Attribution and Repair	Real-World Failure Collection	arXiv	2025	200 trajectories	https://arxiv.org/abs/2509.25370	https://github.com/ulab-uiuc/AgentDebug
TraceElephant	Datasets and Benchmarks for Failure Attribution and Repair	Real-World Failure Collection	ACL	2026	220 trajectories, full observability + reproducible environment	https://arxiv.org/abs/2604.22708	https://github.com/TraceElephant/TraceElephant
Demystifying the Lifecycle of Failures in Platform-Orchestrated Agentic Workflows	Datasets and Benchmarks for Failure Attribution and Repair	Real-World Failure Collection	arXiv	2025	307 trajectories, lifecycle-level annotation + repair strategy	https://arxiv.org/abs/2509.23735v2
AgentRx	Datasets and Benchmarks for Failure Attribution and Repair	Real-World Failure Collection	arXiv	2026	first unrecoverable failure step annotation	https://arxiv.org/abs/2602.02475
CodeTraceBench	Datasets and Benchmarks for Failure Attribution and Repair	Real-World Failure Collection	arXiv	2026	software-engineering agent traces for debugging and patching	https://arxiv.org/abs/2604.11641
MP-Bench	Datasets and Benchmarks for Failure Attribution and Repair	Real-World Failure Collection	arXiv	2026	multi-perspective attribution with failure reason and ideal action annotations	https://arxiv.org/abs/2603.25001
Aegis	Datasets and Benchmarks for Failure Attribution and Repair	Synthetic Data via Error Injection	arXiv	2025	9,533 trajectories	https://arxiv.org/abs/2509.14295
CORRECT-Error	Datasets and Benchmarks for Failure Attribution and Repair	Synthetic Data via Error Injection	arXiv	2025	2,000+ trajectories	https://arxiv.org/abs/2509.24088
Understanding Software Engineering Agents: A Study of Thought-Action-Result Trajectories	Others and Empirical Studies		ASE	2025		https://arxiv.org/abs/2506.18824
MAESTRO: Multi-Agent Evaluation Suite for Testing, Reliability, and Observability	Others and Empirical Studies		arXiv	2026		https://arxiv.org/abs/2601.00481	https://github.com/sands-lab/maestro
Trajectory Guard — A Lightweight, Sequence-Aware Model for Real-Time Anomaly Detection in Agentic AI	Others and Empirical Studies		AAAI Workshop	2026		https://arxiv.org/abs/2601.00516
From Features to Actions: Explainability in Traditional and Agentic AI Systems	Others and Empirical Studies		arXiv	2026		https://arxiv.org/abs/2602.06841			https://vectorinstitute.github.io/unified-xai-evaluation-framework/
Signals: Trajectory Sampling and Triage for Agentic Interactions	Others and Empirical Studies		arXiv	2026		https://arxiv.org/abs/2604.00356
Beyond Final Code: A Process-Oriented Error Analysis of Software Development Agents in Real-World GitHub Scenarios	Others and Empirical Studies		arXiv	2025		https://arxiv.org/abs/2503.12374
Beyond Resolution Rates: Behavioral Drivers of Coding Agent Success and Failure	Others and Empirical Studies		arXiv	2026		https://arxiv.org/abs/2604.02547

A Survey for LLM Agent Trajectory Analysis

This dataset repository hosts the survey paper A Survey for LLM Agent Trajectory Analysis: From Failure Attribution to Enhancement and a structured metadata snapshot of the companion paper collection from Awesome-LLM-Agent-Trajectory-Analysis.

The repository is intended for discovery, citation, and lightweight analysis of the literature around LLM agent trajectory analysis, including failure attribution, trajectory-based debugging, repair, optimization, monitoring tools, and benchmarks.

Note: this is a paper and literature-metadata repository. It is not a raw LLM-agent trajectory corpus.

Path	Description
`LLMAgentTraceAnalysisSurvey.pdf`	The survey paper PDF.
`data/papers.jsonl`	Structured paper/tool/benchmark metadata extracted from the companion awesome-list.
`data/papers.csv`	CSV version of the same metadata for spreadsheet use.
`metadata/awesome_list.md`	Snapshot of the upstream GitHub README used to generate the structured files.
`metadata/paper_collection_summary.json`	Category counts and source metadata for the generated paper list.
`metadata/upstream_LICENSE`	MIT license file from the companion GitHub awesome-list repository.
`assets/cover.png`	Cover image used by the dataset card.
`scripts/build_paper_metadata.py`	Reproducible script for rebuilding `data/papers.*` from `metadata/awesome_list.md`.

Dataset Summary

The survey organizes LLM agent trajectory analysis along five main dimensions:

Failure Taxonomy
Failure Attribution
System Enhancement and Optimization
Trajectory Monitoring and Analysis Tools
Datasets and Benchmarks

The survey reports 55 retained papers from 1,652 initially retrieved papers. The structured files in this repository contain 73 classified records because some works appear in multiple taxonomy roles and the benchmark section includes dataset/benchmark entries in addition to paper entries.

Data Fields

Each row in data/papers.jsonl and data/papers.csv contains:

Field	Description
`title`	Paper, tool, dataset, or benchmark title.
`category`	Top-level taxonomy category from the companion awesome-list.
`subcategory`	Fine-grained taxonomy bucket, when available.
`venue`	Venue or source label parsed from the upstream badge text.
`year`	Publication or listing year parsed from the upstream badge text.
`paper_url`	Paper URL, including arXiv, DOI, ACM DL, or ACL Anthology links.
`code_url`	GitHub repository URL, when listed.
`dataset_url`	Dataset page URL, when listed.
`project_url`	Project or homepage URL, when listed.
`doi_url`	DOI URL, when available.
`other_urls`	Additional URLs that do not fit the fields above.
`note`	Short note from the upstream list, commonly benchmark size or annotation detail.

Category Counts

Category	Records
Failure Taxonomy	9
Failure Attribution Methods for LLM Agents	22
Enhancement, Optimization, and Repair	14
Trajectory Monitoring, Debugging, and Analysis Tools	7
Datasets and Benchmarks for Failure Attribution and Repair	14
Others and Empirical Studies	7

Usage

from datasets import load_dataset

ds = load_dataset(
    "RobinChen2001/A-Survey-for-LLM-Agent-Trajectory-Analysis",
    "papers",
)
print(ds["train"][0])

You can also use the CSV file directly:

import pandas as pd

papers = pd.read_csv("data/papers.csv")
print(papers.groupby("category").size())

Source

Companion GitHub repository: IcyFeather233/Awesome-LLM-Agent-Trajectory-Analysis
Survey PDF in this repository: LLMAgentTraceAnalysisSurvey.pdf
ResearchGate page: A Survey for LLM Agent Trajectory Analysis: From Failure Attribution to Enhancement

Snapshot prepared on 2026-07-09 from the upstream GitHub repository.

Updating the Metadata

After refreshing metadata/awesome_list.md from the companion GitHub repository, rebuild the structured files with:

python3 scripts/build_paper_metadata.py

This regenerates:

data/papers.jsonl
data/papers.csv
metadata/paper_collection_summary.json

Citation

@article{wang2026surveytrajectory,
  title={A Survey for LLM Agent Trajectory Analysis: From Failure Attribution to Enhancement},
  author={Wang, Junjie and Wang, Yawen and Chen, Mengzhuo and Xie, Xiaofei and Chen, Chunyang and Mu, Fangwen and Liu, Zhe and Wang, Qing},
  year={2026}
}

Contact

For questions, suggestions, or collaboration opportunities:

Junjie Wang: junjie@iscas.ac.cn
Mengzhuo Chen: chenmengzhuo2023@iscas.ac.cn

License

The dataset card metadata declares cc-by-4.0 for this Hugging Face repository. The companion awesome-list repository is released under MIT; its license snapshot is included at metadata/upstream_LICENSE. External papers, code repositories, datasets, and project pages linked from the metadata retain their own licenses.

Downloads last month: 246

Number of rows:

Total file size:

1.39 MB