---
pretty_name: StreamStop
language:
- en
license: cc-by-4.0
viewer: false
size_categories:
- 1M<n<10M
tags:
- language-model-agents
- adaptive-stopping
- reasoning
- offline-evaluation
- arxiv:2610.27981
---
# StreamStop: Agent Reasoning Trajectories with Prefix-Level Action Annotations

**JunHyeok Oh and Byung-Jun Lee, Korea University**  
[Paper](https://arxiv.org/abs/2610.27981) · [Project page](https://jh2525.github.io/StreamStop/)

Dataset accompanying *Adaptive Stopping of Streamed Reasoning for Language Model Agents*.

A completed reasoning trajectory records its final action, but does not reveal the actions the model would generate from earlier reasoning prefixes. StreamStop supplies these prefix-action annotations while holding each decision's environment context fixed. Researchers can train stopping policies and compare their action agreement and reasoning-token use offline, without additional reasoning-model calls or environment rollouts after collection.

## Collection

| Configuration | Episodes | Nonempty reasoning streams | Prefix-action annotations |
|---|---:|---:|---:|
| WebShop / Qwen3.8-27B | 5,237 | 40,922 | 443,611 |
| WebShop / Nemotron-30B | 3,437 | 20,021 | 374,499 |
| ScienceWorld / Qwen3.8-27B | 1,260 | 30,885 | 791,006 |
| ScienceWorld / Gemini 3.1 Flash-Lite | 1,260 | 31,991 | 487,700 |
| **Total** | **11,194** | **123,819** | **2,096,816** |

The full collection includes successful and unsuccessful episodes and invalid action responses. It contains 123,820 decision records, including one Gemini decision with empty reasoning. Annotation coverage varies across streams. The paper's training and validation experiments use a filtered subset from successful collection episodes; the full collection should not be confused with that subset. Closed-loop test evaluation rollouts are not included.

Gemini uses an explicit plan with native thinking disabled. Its prefix lengths use the prefix reader's tokenizer units; they do not represent native reasoning traces or API billing savings.

## Files and records

Each of `webshop_q38`, `webshop_nemo`, `sciworld_q38`, and `gemini_sciworld` contains:

- `episodes.jsonl.gz`: episode summaries, task identifiers, success and outcomes.
- `decisions.jsonl.xz`: public context, full reasoning, reference action and parser results.
- `prefix_actions.jsonl.xz`: prefix length `k`, generated responses/actions and reference parser results.

This release provides the original compressed JSONL files. The automatic viewer is disabled because episodes, decisions, and prefix annotations have different record schemas; use the download and reading example below.

All are UTF-8 JSON Lines. JSON records retain their original fields; optional fields differ across configurations. `manifest.json` records collection counts and checksums.

Join decisions and annotations using configuration, `source_shard`, `split`, `goal`, `step`, and `attempt` (default zero when absent). Episodes use configuration, `source_shard`, `split`, and `goal`. `k` and `sample_index` distinguish prefix annotations. Keep repeated collection samples separate, and preserve the recorded task-level split when building experiments.

## Download and read

```python
from huggingface_hub import snapshot_download
from pathlib import Path
import gzip, json, lzma

root = Path(snapshot_download("jh2525/StreamStop", repo_type="dataset"))

def iter_records(path):
    opener = lzma.open if path.suffix == ".xz" else gzip.open
    with opener(path, "rt", encoding="utf-8") as stream:
        for line in stream:
            if line.strip():
                yield json.loads(line)

prefixes = iter_records(root / "webshop_q38/prefix_actions.jsonl.xz")
record = next(prefixes)
print(record["k"], record["parse"])
```

Compressed files can be read sequentially without decompressing the complete collection to disk. `examples/read_records.py` supplies a local reading example.

## Interpretation

Action agreement compares a prefix action with the recorded full-reasoning reference action. It is not a ground-truth correctness label: another action can also succeed in the environment. Offline agreement evaluation is distinct from closed-loop task-success evaluation. See the paper for collection prompts, parsing, filtering, and stopping protocols.

## License and source terms

**CC BY 4.0 applies to our original prefix-action annotations.** Benchmark content and model outputs retain their applicable upstream terms. This license does not relicense those assets. See [SOURCE_TERMS.md](SOURCE_TERMS.md) and [LICENSE](LICENSE).

No model weights are included. Provider request identifiers, credentials, and internal filesystem paths are not part of this release.

## Citation

Please cite the accompanying [paper](https://arxiv.org/abs/2610.27981):

```bibtex
@article{oh2026streamstop,
  title={Adaptive Stopping of Streamed Reasoning for Language Model Agents},
  author={Oh, JunHyeok and Lee, Byung-Jun},
  journal={arXiv preprint arXiv:2610.27981},
  year={2026},
  eprint={2610.27981},
  archivePrefix={arXiv},
  primaryClass={cs.CL},
  url={https://arxiv.org/abs/2610.27981}
}
```