| --- |
| language: |
| - en |
| license: cc-by-4.0 |
| tags: |
| - benchmark |
| - agents |
| - evaluation |
| - synthetic |
| pretty_name: DMAIC-Bench |
| --- |
| |
| # DMAIC-Bench |
|
|
| [Paper: arXiv:2610.32305](https://arxiv.org/abs/2610.32305) · [Code and evaluation framework](https://github.com/psptpx/DMAIC-Bench) |
|
|
| DMAIC-Bench evaluates whether agents use intermediate evidence to continue, divert, or stop in multi-stage analytical workflows. The paper describes 97 seeded tasks across 15 industrial cases and a 16-task life-sciences extension. |
|
|
| ## Scope and usage |
|
|
| The benchmark is an interactive evaluation: generated data is released in stages by a steward, and a separate verifier evaluates execution. Browsing or loading dataset rows alone does not reproduce the full agent evaluation. Refer to the GitHub repository for task generation, Harbor integration, scoring, dependencies, and instructions. |
|
|
| This public export contains 78 records: 65 clean tasks across 10 public cases and 13 fault variants, generated using the public-only build of the GitHub release. Each JSONL row includes the task ID, case, industry, seed, fault flag, agent instruction, and synthetic data blocks as a JSON string. All blocks are visible in this export; access restrictions and the steward are not enforced by this dataset viewer. Reference solutions, verifier truth and runtime code are not included here. See the GitHub repository to construct and run the full environment. The life-sciences extension is described in the paper but is not included in this export. The original industrial campaign per-trial archive is not included in the public GitHub release. Reported experiments covered all 97 industrial tasks before publication; do not describe those results as a held-out generalization evaluation. |
|
|
| ## License |
|
|
| The generated benchmark corpus is distributed under CC BY 4.0. Software in the GitHub repository is separately licensed under AGPL-3.0-only. The paper is separately licensed under CC BY 4.0. |
|
|
| ## Citation |
|
|
| ```bibtex |
| @misc{peng2026dmaicbench, |
| title={DMAIC-Bench: From Self-Reflection to Self-Restraint for Agents in Multi-Stage Analytical Tasks}, |
| author={Peng, Shan and Cheng, Shuaiqi and Peng, Miao and Zi, Chenyi and Li, Jia}, |
| year={2026}, |
| eprint={2610.32305}, |
| archivePrefix={arXiv}, |
| primaryClass={cs.AI}, |
| url={https://arxiv.org/abs/2610.32305} |
| } |
| ``` |
|
|
