DreamEternal/SelfImprove-Bench-Cases

SelfImprove-Bench: Evaluating Generalization in Self-Improving Agents

✨ Introduction

This dataset contains instance cases of SelfImprove-Bench. They are instances generated with different task templates, difficulty coefficients, and random seed. The full code of the benchmark is available at https://github.com/theseus-labs-rsi/SelfImprove-Bench. Its protocol supports multiple representative RSI methods to self-evolve from feedback on multi-round tasks, and provides task generation, simulation, and evaluation through six game scenarios across three capability categories. With a scalable task generation and orchestration mechanism, the benchmark can measure the effectiveness of RSI methods at different levels of generalization and offers better sustainability.

• Generalization Ladder: Organize evaluation into generalization ladder levels G0 to G4: same instances (G0), unseen instances at the same difficulty (G1), new difficulty levels within the same task (G2), different tasks within the same capability category (G3), and tasks across capability categories (G4).
• Scalable Task Generation and Orchestration: Automatic generate task istances based on templates and difficuly coefficients, enabling configurable task sequences for evaluating each level of the generalization ladder.
• High Resolution Trace Analysis: Characterize refinement semantics, modification scope, capability dimension, and generalizability from evolution trajectories, enabling a more comprehensive reflection on the method pattern.

The initial version covers following methods and games, while you can extend new ones based on the extensible framework.

• Methods: RSI framework (ACE, AHE, HyperAgents) and baselines (Codex, Claude Code, Deepseek Harness).
• Games: Farm, Combat, Sweepmine, Hero's Journey, Minecraft, Astray.

📚 Citation

If you find this project useful for your research, please consider giving us a star and citing our paper.

@article{xu2026selfimprovebench,
  title={SelfImprove-Bench: Evaluating Generalization in Self-Improving Agents},
  author={Bangrui Xu and Jie Yang and Sidi Chen and Xuanhe Zhou and Jiantao Qiu and Qianyu Yang and Zijie Wang and Fan Wu and Zhoufutu Wen},
  journal={arXiv preprint arXiv:2611.02797},
  year={2026}

📄 License

This project is licensed under the MIT License. See the LICENSE file for details. Bundled third-party components retain their own licenses:

Component	License / terms
ACE	Apache-2.0; see modifications
AHE	MIT
NexAU	Apache-2.0; see bundled distribution details
HyperAgents	CC BY-NC-SA 4.0, non-commercial use only
Hero's Journey	MIT
GameWorld runtime and games	Adapted from GameWorld; bundled game assets are for research and educational use only. See attribution and terms.

The repository's MIT License does not override third-party terms.