| --- |
| language: |
| - en |
| pretty_name: AdaExploreBench |
| size_categories: |
| - 1K<n<10K |
| tags: |
| - benchmark |
| - llm-agents |
| - exploration |
| - adaptation |
| - rule-induction |
| configs: |
| - config_name: default |
| data_files: |
| - split: train |
| path: data/train.jsonl |
| - split: test |
| path: data/test.jsonl |
| --- |
| |
| # AdaExploreBench |
|
|
| AdaExploreBench evaluates proactive and adaptive exploration in LLM agents. Agents design probe cases, infer hidden decision rules from utility feedback, and adapt to unannounced environmental changes. |
|
|
| [Code](https://github.com/HKUST-KnowComp/AdaExploreBench) · [Eight-model overview](https://hkust-knowcomp.github.io/AdaExploreBench/) |
|
|
| ## Splits |
|
|
| | Split | Episodes | Availability | |
| | --- | ---: | --- | |
| | `test` | 256 | Available in this release | |
| | `train` | 1,024 | Available in this release | |
|
|
| Both splits are balanced across finance, healthcare, cybersecurity, and manufacturing, and across additive aggregation, conjunctive gating, disjunctive sufficiency, and lexicographic priority rules. Each dynamic episode has eight numeric fields, three active factors per phase, and four ordered actions. An unannounced change replaces one factor and reverses another factor's direction. |
|
|
| ## Files |
|
|
| - `data/train.jsonl` contains 1,024 training episodes, one per line. |
| - `data/test.jsonl` contains 256 test episodes, one per line. |
| - `manifest.json` records the split size, schema version, domain and rule composition, and SHA-256 checksum. |
|
|
| Episodes contain the evaluator's full task specification, latent mechanisms, evaluation cases, and branch definitions. Evaluated agents receive the task and feedback through the benchmark runtime, which controls access to hidden information. |
|
|
| Use the code repository's `scripts/run_benchmark.py` with the downloaded JSONL path. Test episodes are intended for evaluation; keep them separate from model training and prompt development. |
|
|
| ## Train/test separation |
|
|
| The training and test splits have disjoint episode IDs and scenario template IDs. The training split covers 32 scenario templates and the test split covers 16. Use the training split for model training and development, and reserve the test split for evaluation. |
|
|
| ## Loading |
|
|
| Download the JSONL files for use with the benchmark runtime: |
|
|
| ```bash |
| hf download Roryaccout/AdaExploreBench --repo-type dataset --local-dir datasets/adaexplorebench |
| ``` |
|
|
