Continual Learning Exam: Benchmarking Long-Horizon RSI and Experiential Learning in LLM Agents
Paper • 2610.33956 • Published • 24
Request access to the CLE Lite full packages
This dataset holds the complete CLE Lite task packages, including the evaluation rubrics, ground truth, grader and reference solutions. Access is gated to protect the benchmark and is approved manually.
Log in or Sign Up to review the conditions and access this dataset content.
CLE Lite — Full
The complete packages of the twenty CLE Lite tasks, for running and scoring agents.
The complete packages of the twenty CLE Lite tasks are available here after manual access approval. Each task directory holds one Harbor package and the notices for the third-party files it contains:
| Path | Contents |
|---|---|
harbor/task/ |
The Harbor task: container environment, workspace files, the five level instructions, tests and the reference-solution entry point |
harbor/private/ |
The judge side: rubric checks, reference values and deliverables, the reference program and the sealed grader |
harbor/build.json, harbor/freeze.json |
The build record and the SHA-256 manifest of every task and judge file |
NOTICE.md, PROVENANCE.md |
Third-party sources and terms; the source, version, retrieval time, hash and licence note of every workspace file |
The runtime/ directory holds the CLE Lite runner; runtime/README.md explains how to download a task with its integrity checks, run an agent in direct or CLE mode, and grade the submissions.
Use this material only in an evaluation setup that keeps judge-side files separate from the agent.
🗂️ The four CLE Lite repositories
| Repository | Contents | Access |
|---|---|---|
| CLE Lite | 100 task cards: assignments, expected inputs and outputs, input-file inventories | Public |
| CLE Lite Workspaces | Level instructions and the workspace files supplied to agents | Public |
| CLE Lite Rubrics | Evaluation rubrics and ground-truth references | Approval required |
| CLE Lite Full (this repository) | Complete task packages for running and scoring agents | Approval required |
📁 Layout
016-data-scientist/
├── NOTICE.md
├── PROVENANCE.md
└── harbor/
├── task/
│ ├── task.toml instruction.md
│ ├── steps/t1/ … steps/t5/
│ ├── environment/
│ ├── tests/
│ └── solution/
├── private/
│ ├── checks.json task.json
│ ├── reference/
│ └── grader/
├── build.json
└── freeze.json
runtime/ the CLE Lite runner and its tests
The Hub does not store folders named .git. Task 009-localization-engineer holds real git checkouts; their .git folders are in harbor/git-folders.tar. The runtime's download command unpacks and verifies them; after a plain hf download, run tar -xf git-folders.tar inside harbor/.
⚠️ Terms of use
For open inspection, see the public CLE Lite cards and CLE Lite Workspaces. These public resources omit judge-side answers and setup scripts.
Do not include the tasks or reference material in training corpora.
📝 Citation
If you use these CLE Lite resources, please cite the CLE paper:
@misc{zhao2026continuallearningexambenchmarking,
title = {Continual Learning Exam: Benchmarking Long-Horizon RSI and Experiential Learning in LLM Agents},
author = {Haoyu Zhao and Sitian Tu and Yuxuan Huang and Huichi Zhou
and Zhangyue Yin and Jiahao Huang and Jingyuan Xu and Tianmiao Shi
and Jin Shi and Nianxi Zhang and Yangmiao Zhang and Zhuolang Chan
and Huiyan Hu and Chenchao Sheng and Yaqing Luo and Yue Chen
and Huaxiu Yao and Meng Fang and Jun Wang},
year = {2026},
eprint = {2610.33956},
archivePrefix = {arXiv},
primaryClass = {cs.LG},
url = {https://arxiv.org/abs/2610.33956}
}
- Downloads last month
- 101