LongRCA Bench: Diagnosing Responsible Roles and Root Causes in Long-Horizon Agent Failures
Paper • 2608.15242 • Published • 15
question_ID stringlengths 14 22 | history listlengths 10 728 | mistake_agent stringclasses 13
values | mistake_step int64 1 521 | mistake_reason stringlengths 66 5.4k |
|---|---|---|---|---|
swe_bench_pro__001 | [{"step":0,"content":"You are working on the repository \"ansible/ansible\" located at /app.\n\n## P(...TRUNCATED) | DiagnostAgent (-> ActionAgent) | 25 | "Final eval came back 35/39 with reward=0. The four failing tests are the three TestParseContent cas(...TRUNCATED) |
swe_bench_pro__002 | [{"step":0,"content":"You are working on the repository \"NodeBB/NodeBB\" located at /app.\n\n## Pro(...TRUNCATED) | DiagnostAgent (-> ActionAgent) | 21 | "The root cause should be labeled as step 21. In this step, when DiagnostAgent delegated to ActionAg(...TRUNCATED) |
swe_bench_pro__003 | [{"step":0,"content":"You are working on the repository \"NodeBB/NodeBB\" located at /app.\n\n## Pro(...TRUNCATED) | DiagnostAgent (-> ActionAgent) | 246 | "The repair plan issued by DiagnostAgent through delegate_to_action itself had flaws - it only liste(...TRUNCATED) |
swe_bench_pro__004 | [{"step":0,"content":"You are working on the repository \"NodeBB/NodeBB\" located at /app.\n\n## Pro(...TRUNCATED) | DiagnostAgent (-> ActionAgent) | 124 | "verification_gap \nAt step 124, DiagnostAgent incorrectly instructed ActionAgent to replace `db.del(...TRUNCATED) |
swe_bench_pro__005 | [{"step":0,"content":"You are working on the repository \"NodeBB/NodeBB\" located at /app.\n\n## Pro(...TRUNCATED) | DiagnostAgent (-> ActionAgent) | 33 | "The only failed required test is `user.uploadCroppedPicture > should remove uploaded picture` (310/(...TRUNCATED) |
swe_bench_pro__006 | [{"step":0,"content":"You are working on the repository \"NodeBB/NodeBB\" located at /app.\n\n## Pro(...TRUNCATED) | ActionAgent | 56 | "deleteFromDisk threw the generic error [[error:invalid-data]] for non-string, non-array input param(...TRUNCATED) |
swe_bench_pro__007 | [{"step":0,"content":"You are working on the repository \"NodeBB/NodeBB\" located at /app.\n\n## Pro(...TRUNCATED) | DiagnostAgent (-> ActionAgent) | 27 | "DiagnostAgent, in the delegated plan, required replacing the original isValidationPending check in (...TRUNCATED) |
swe_bench_pro__008 | [{"step":0,"content":"You are working on the repository \"NodeBB/NodeBB\" located at /app.\n\n## Pro(...TRUNCATED) | ActionAgent | 54 | "Final eval came back 3509/3510 with reward=0, and the single casualty is test/messaging.js `.canMes(...TRUNCATED) |
swe_bench_pro__009 | [{"step":0,"content":"You are working on the repository \"NodeBB/NodeBB\" located at /app.\n\n## Pro(...TRUNCATED) | ActionAgent | 52 | "While modifying the `deleteInvitationKey` function, ActionAgent failed to clean up the `invitation:(...TRUNCATED) |
swe_bench_pro__010 | [{"step":0,"content":"You are working on the repository \"NodeBB/NodeBB\" located at /app.\n\n## Pro(...TRUNCATED) | ActionAgent | 76 | "When implementing Topics.syncBacklinks, ActionAgent misjudged the usage of NodeBB's Redis sorted se(...TRUNCATED) |
End of preview. Expand in Data Studio
LongRCA Bench
A benchmark for diagnosing responsible roles and root causes in long-horizon agent failures.
- 1,140 observed, non-injected failed trajectories from SWE-bench Pro, Terminal-Bench 2, TravelPlanner, VitaBench, and WebArena Verified.
- Human annotations identifying the responsible role, earliest decisive root-cause step, and supporting rationale.
- LongRCA-Mini: a fixed 200-trajectory subset, randomly sampled without replacement with 40 from each benchmark, preserving the original IDs and annotations.
Individual JSON files are grouped by benchmark in data/full/ and data/mini/. The corresponding Parquet files are available at the repository root.
See the paper and project page for details. If you use either version, please cite our paper.
Usage
from datasets import load_dataset
full = load_dataset("CLoud5-real/longrca-bench", split="test")
mini = load_dataset("CLoud5-real/longrca-bench", "mini", split="test")
- Downloads last month
- 469