README.md

---
license: cc-by-4.0
pretty_name: RL Method Tracker
language:
- en
size_categories:
- 1K<n<10K
tags:
- reinforcement-learning
- llm
- post-training
- rlhf
- open-source-software
- mining-software-repositories
configs:
- config_name: papers
  data_files: papers.parquet
  default: true
- config_name: methods
  data_files: methods.parquet
- config_name: landings
  data_files: landings.parquet
- config_name: proposals
  data_files: proposals.parquet
---

# RL Method Tracker

Which RL post-training methods for language models reached seven open-source training frameworks (verl, TRL, OpenRLHF, NeMo-RL, ROLL, slime, AReaL), when they landed, and how each framework offers them.
Of 1,598 arXiv papers from January 2025 to June 2026 that propose a general RL training method, 37 reached the main branch of at least one framework.

- Paper: [From RL Papers to Framework Code: Inclusion and Review in Seven Open-Source Training Frameworks](https://arxiv.org/abs/2610.28027)
- Project page with search: https://pth2002.github.io/rl-method-tracker/
- Code and the full data release: https://github.com/pth2002/rl-method-tracker
- Archived release: [doi:10.5281/zenodo.23495744](https://doi.org/10.5281/zenodo.23495744)

This is the frozen v1.0 data behind the report, with framework histories up to the 2026-10-18 snapshot.
The GitHub repository refreshes the framework histories every week.

## Load

```python
from datasets import load_dataset

papers = load_dataset("Purdy0228/rl-method-tracker", "papers", split="train")
methods = load_dataset("Purdy0228/rl-method-tracker", "methods", split="train")
```

## Subsets

| Subset | Rows | One row per |
|---|---|---|
| `papers` | 6,939 | screened arXiv paper on RL for language models |
| `methods` | 65 | method in the catalog, with its first landing in each framework |
| `landings` | 101 | first commit that put a method on a framework's main branch |
| `proposals` | 48 | public proposal for a general method that no framework included |

### papers

- `arxiv`, `title`, `month`: arXiv id, title and month of the first version.
- `label`: what the paper contributes. `rl_method`, `rl_application`, `analysis`, `systems`, `distillation`, `preference_offline` or `other`.
- `scope`: for RL methods, `general` when the method is the paper's main contribution and `embedded` when it sits inside a paper about a task, domain, agent or application.
- `component`: the part of training the method changes, such as `advantage`, `clipping`, `sampling` or `entropy`.
- `label_confidence`: `high`, `medium` or `low`.
- `is_model_report`: the paper is the technical report of a model.
- `general_method`: the paper counts toward the 1,598 general methods.
- `included`: the paper's method reached the main branch of at least one framework.
- `methods`: catalog ids of the included methods the paper introduces.
- `n_frameworks_ever`, `n_frameworks_at_snapshot`: frameworks that ever included the method, and those that still include it at the snapshot.

### methods

- `id`, `name`, `family`, `arxiv`, `source_date`: the method and its first public source.
- `n_fw_ever`, `n_fw_head`, `n_default`: frameworks that ever included it, that include it at the snapshot, and that turn it on by default.
- `first_adoption`, `first_lag_days`, `first_lag_is_upper_bound`: date of the first landing in any framework and the days from the source to it. The lag is an upper bound when that first landing arrived with a repository's initial commit or a bulk release.
- `n_self_contributed`: frameworks whose landing commit came from an author of the paper.
- `<framework>_first`: date of the first landing in that framework.
- `<framework>_head`: tier at the snapshot. `default` is on in a run that keeps every setting. `conditional_default` is on once a related feature is enabled. `core` is a selectable algorithm in the main code path. `option` is a switch that modifies the main algorithm. `experimental` and `recipe` sit outside the main path.

### landings

- `method`, `framework`, `date`, `pr`, `tier`: the landing and the pull request that made it.
- `first_in_any`: this framework was the first to include the method.
- `form`: how the method changes training, as `coefficient`, `term` or `pipeline`.
- `commit_author_relation`: whether the landing commit came from an author of the paper.
- `prior_request`: an issue or pull request asked for the method before it landed.
- `pr_author`, `merged_by`, `author_prior_merger`, `self_merged`, `human_reviewers`: who proposed and merged the pull request. A prior merger had merged a pull request in the framework before.
- `days_open`, `core_lines`, `files`, `lines`, `tests`, `docs`, `deps`, `new_files`: size and contents of the change.

### proposals

- `arxiv`, `method`, `framework`, `type`, `number`: the proposal, as a pull request (`pr`) or a request in an issue or discussion (`request`).
- `outcome`: `open_no_reply`, `open_engaged`, `declined`, `redirected`, `stale_cleanup`, `closed_inactive`, `closed_no_comment` or `withdrawn`.
- `reasons`: concerns the maintainers stated, such as `demand`, `surface`, `placement` or `evidence`.
- `quote`: one short quote from the thread as evidence for the code.

Thread texts stay on GitHub. The labelling rules and coding guides are in the [repository](https://github.com/pth2002/rl-method-tracker).

## Citation

```bibtex
@misc{pan2026rlframeworks,
  title         = {From {RL} Papers to Framework Code: Inclusion and Review in Seven Open-Source Training Frameworks},
  author        = {Pan, Taiheng},
  year          = {2026},
  eprint        = {2610.28027},
  archivePrefix = {arXiv},
  primaryClass  = {cs.LG},
  doi           = {10.48550/arXiv.2610.28027},
  url           = {https://arxiv.org/abs/2610.28027}
}
```

## License

CC BY 4.0. arXiv titles come from arXiv metadata.