README.md

---
pretty_name: "LongRCA Bench"
language:
- en
size_categories:
- 1K<n<10K
annotations_creators:
- expert-generated
tags:
- agents
- multi-agent-systems
- failure-analysis
- root-cause-analysis
configs:
- config_name: default
  default: true
  data_files:
  - split: test
    path: "data/test-*.parquet"
---

# LongRCA Bench

LongRCA Bench contains 1,140 observed, non-injected failed agent
trajectories from five task domains. Each trajectory has human annotations for
the responsible role, earliest decisive root-cause step, and a
trajectory-grounded rationale.

For the benchmark definition, annotation protocol, and evaluation results, see:

**[LongRCA Bench: Diagnosing Responsible Roles and Root Causes in Long-Horizon Agent Failures](https://arxiv.org/abs/2608.15242)**

## Dataset Composition

| Source | Instances |
|---|---:|
| SWE-bench Pro | 128 |
| Terminal Bench 2 | 42 |
| TravelPlanner | 685 |
| VitaBench | 108 |
| WebArena Verified | 177 |
| **Total** | **1,140** |

The dataset contains 178,137 recorded history steps. The median trajectory
length is 145 steps, and the maximum is 728 steps.

## Data Fields

| Field | Description |
|---|---|
| `question_ID` | Formal source-prefixed trajectory identifier (`<source>__NNN`) |
| `history` | Complete ordered trajectory |
| `mistake_agent` | Reference responsible role |
| `mistake_step` | Reference 0-based root-cause step |
| `mistake_reason` | Human-written rationale; not scored |

The responsible role and root-cause step are evaluated independently. The role
must not be inferred automatically from the emitter of `mistake_step`.
Release IDs are numbered independently within each source and should be treated
as stable, opaque keys.

## Usage

```python
from datasets import load_dataset

dataset = load_dataset(
    "CLoud5-real/longrca-bench",
    split="test",
)
```

## Source Composition

| Source benchmark | Task domain | Trajectories |
|---|---|---:|
| [SWE-bench Pro](https://arxiv.org/abs/2509.16941) | Software repair | 128 |
| [Terminal Bench 2](https://github.com/harbor-framework/terminal-bench-2) | Terminal tasks | 42 |
| [TravelPlanner](https://github.com/OSU-NLP-Group/TravelPlanner) | Travel planning | 685 |
| [VitaBench](https://vitabench.github.io/) | Service-oriented tool use | 108 |
| [WebArena Verified](https://github.com/ServiceNow/webarena-verified) | Web interaction | 177 |
| **Total** | **Five task domains** | **1,140** |

The trajectories were generated with MiniMax-M2.5, Kimi-K2.5, and
Qwen3.5-Plus under several agent organizations, including fixed-role teams,
specialist group chats, sequential workflows, and planner–critic–executor
workflows.

## Citation

If you use LongRCA Bench, please cite:

```bibtex
@misc{zhang2026longrcabench,
  title        = {LongRCA Bench: Diagnosing Responsible Roles and Root Causes
                  in Long-Horizon Agent Failures},
  author       = {Yunfei Zhang and Boyu Feng and Changhua Pei and
                  Zexin Wang and Zhihuang Peng and Xinlong Liu and
                  Hengyue Jiang and Difeng Ma and Jiayi Zhang and
                  Yongzhou Yao and Yanan Zhao and Fei Sun and
                  Yintong Huo and Zhaoyang Liu and Jingjing Li and
                  Gaogang Xie and Dan Pei},
  year         = {2026},
  eprint       = {2608.15242},
  archivePrefix = {arXiv},
  primaryClass = {cs.AI},
  url          = {https://arxiv.org/abs/2608.15242}
}
```