[algo] Add SPO (Single-stream Policy Optimization) recipe implementation

[dzh19990407]

What does this PR do?

Add concise overview of what this PR aims to achieve or accomplish. Reference related GitHub issues and PRs that help with the review.

This PR implements the Single-stream Policy Optimization proposed by paper https://arxiv.org/abs/2509.13232.

Checklist Before Starting

• Search for similar PRs. Paste at least one query link here: [algo] feat: add GSPO-token policy loss computation function #2775
• Format the PR title as [{modules}] {type}: {description} (This will be checked by the CI)
  • {modules} include fsdp, megatron, sglang, vllm, rollout, trainer, ci, training_utils, recipe, hardware, deployment, ray, worker, single_controller, misc, perf, model, algo, env, tool, ckpt, doc, data
  • If this PR involves multiple modules, separate them with , like [megatron, fsdp, doc]
  • {type} is in feat, fix, refactor, chore, test
  • If this PR breaks any API (CLI arguments, config, function signature, etc.), add [BREAKING] to the beginning of the title.
  • Example: [BREAKING][fsdp, megatron] feat: dynamic batching

Test

For changes that can not be tested by CI (e.g., algorithm implementation, new model support), validate by experiment(s) and show results like training curve plots, evaluation results, etc.

API and Usage Example

Demonstrate how the API changes if any, and provide usage example(s) if possible.

# Add code snippet or script demonstrating how to use this

Design & Code Changes

Demonstrate the high-level design if this PR is complex, and list the specific changes.

Checklist Before Submitting

Important

Please check all the following items before requesting a review, otherwise the reviewer might deprioritize this PR for review.

• Read the Contribute Guide.
• Apply pre-commit checks: pre-commit install && pre-commit run --all-files --show-diff-on-failure --color=always
• Add / Update the documentation.
• Add unit or end-to-end test(s) to the CI workflow to cover all the code. If not feasible, explain why: ...
• Once your PR is ready for CI, send a message in the ci-request channel in the verl Slack workspace. (If not accessible, please try the Feishu group (飞书群).)

[CLAassistant]

All committers have signed the CLA.

[gemini-code-assist]

Code Review

This pull request introduces the Single-stream Policy Optimization (SPO) algorithm, a novel reinforcement learning method for Large Language Models. The changes primarily consist of new files for the SPO recipe, including configuration, the main training script, a run script, and the core Ray trainer implementation. My review has identified two critical issues. First, the run_spo.sh script uses an undefined variable which will cause the training to fail at launch. Second, the spo_ray_trainer.py contains unsafe exception handling during data resampling, which could lead to silent data corruption and hard-to-debug training failures. Addressing these issues is crucial for the correctness and stability of the new algorithm's implementation.

[vermouth1992]

Could you pin the verl commit in your readme?

[hustnn]

what is the purpose of this file?

[zhongwen-xu]

see Appendix A

[dzh19990407]

This file is an offline value estimate (Appendix A), I have added a link to huggingface in the README.

[dzh19990407]

Could you pin the verl commit in your readme?

I have update it in the readme file.

[dzh19990407]

@wuxibin89 @vermouth1992 @tongyx361 @PeterSH6
Hi team,

Thanks for all the great feedback! I have updated the code based on the review comments and pushed the changes.

Please take a quick look when you have a chance. Thanks!

[zhongwen-xu]

@hustnn @wuxibin89 @vermouth1992 @tongyx361 @PeterSH6

Hey folks, can you review this PR?

[hustnn]

Hi @zhongwen-xu ，I am not in the verl team. I am interested in your work and testing it in my case (a coding model). If it working, I will let you know.

Do I need to first run it to get offline values before training?

[zhongwen-xu]

Hi @hustnn,

Thanks for your interests. For your question, please refer to Appendix E in https://arxiv.org/abs/2509.13232, especially Figure 6 (c) for "Offline Initialization Ablation".

Short answer: the offline initialization helps for early training steps, in the long run, with and without the initializaon are similar.

My recommendation is to get the offline values before training though, a good v_hat estimation with the offline sampling is relatively cheap, same cost as what people do for filtering prompts, and they can be shared across multiple experiment runs.

Happy to answer any further questions by email!

[zsgvivo]

Hi @zhongwen-xu ,Thanks for your great work on SPO. I'm testing your implementation using the training scripts provided in this PR (recipe/spo/run_spo.sh). I've noticed that in both SPO and GRPO setup, the model use tools at the beginning, but as training progresses, it gradually stops calling them and instead generates direct answers. Is this behavior expected?

[zhongwen-xu]

Hi @zsgvivo

Thanks for your interests in our work!
The reason we decided to close this PR was that we initially thought a quick PR with the original ReTool tool protocol might work with Qwen3 8B without SFT, so we submitted the PR before the holiday. However, during the long holiday, we did a verification run with this PR implementation, it turned out it did't work, similar to what you just reported. Now we suspected that the ReTool protocol requires SFT first, rather than directly RL from it.

We are working on a full replication of what we had in the paper where the tool call protocol is with
```python
the generated code here
```

, rather than the JSON function calls as it was in ReTool implementation. (Note that we actually present all the core code in this PR so if you can modify the tool call protocol as what I just said, it should just work, without waiting for our implementation.)

Stay tuned!

[longls777]

Hi @zhongwen-xu and SPO authors,

Thanks a lot for sharing your work and the initial SPO implementation in this PR.

From your last comment, I understand that this PR version didn’t pass your verification run and that you’re working on a full replication with the updated tool-calling protocol (instead of the original ReTool JSON function calls). I’m very interested in reproducing the SPO results and applying the method to my own use cases.

May I ask if there is currently a validated/recommended implementation of SPO available (e.g., a newer branch, updated scripts, or a different repo) that you would suggest users follow?
If there is any WIP code, configs, or guidance you’re comfortable sharing—even if it’s not fully polished yet—it would be extremely helpful.

Thanks again for the great work, and really appreciate any pointers!

[dzh19990407]

@longls777 SPO has been merged into https://github.com/verl-project/verl-recipe

[longls777]

Thanks so much for sharing this implementation - really appreciate your help!