RedundancyBench

Paper · Code and evaluation · Data files

RedundancyBench contains 1,920 agent trajectories and 3,613 step-level redundancy annotations across Airline, Retail, and Telecom.

• human_annotation: 189 human-annotated trajectories with 657 labeled steps.
• automated_annotation: 1,731 trajectories with 2,956 model-consensus labels; these trajectories were not individually human-verified.
• Annotations identify the redundant step, its type, supporting evidence, and a natural-language reason.
• Four types: Duplicated Step, Abnormal Step, Incorrect Step, and Exploratory Step.

Both subsets use a single test split. All released trajectories contain at least one labeled redundant step. Task definitions, annotation details, and evaluation instructions are available in the GitHub repository.

Usage

Each row contains one trajectory. Decode messages_json and annotation_json to access its messages and full annotation.

import json
from datasets import load_dataset

data = load_dataset(
    "redundant-or-necessary/RedundancyBench",
    "human_annotation",  # or "automated_annotation"
    split="test",
)
messages = json.loads(data[0]["messages_json"])
annotation = json.loads(data[0]["annotation_json"])

Reference

If you use RedundancyBench, please cite:

@article{zhu2026redundant,
  title   = {Redundant or Necessary? A Benchmark for Detecting Redundant Steps in Agent Trajectories},
  author  = {Zhu, Liang and Yang, Bo and Zhou, Zhinuo and Liang, Jiachen and
             Guo, Jiahao and Yin, Yiyang and Han, Xiongwei and Hu, Minyang},
  journal = {arXiv preprint arXiv:2605.29893},
  year    = {2026}
}