---
pretty_name: "SSPO: Step-Level Self-Distilled Policy Optimization Data"
language:
  - en
license: mit
task_categories:
  - question-answering
  - text-generation
tags:
  - deep-search
  - web-search
  - web-agents
  - reinforcement-learning
  - rlvr
  - reasoning
  - tool-use
  - sft
  - trajectories
configs:
  - config_name: evidence_anchors
    data_files:
      - split: train
        path: evidence_anchors.jsonl
  - config_name: gpt_oss_120b_trajectories
    data_files:
      - split: train
        path: gpt-oss-120b-traj.jsonl
---

<div align="center">
  <h1 align="center">SSPO</h1>
  <h3 align="center">Beyond Outcome Rewards: Step-Level Self-Distilled Policy Optimization for Deep Search Agents</h3>

  <p align="center">
    <a href="#">📄 arXiv</a> •
    <a href="https://github.com/hkust-nlp/SSPO">💻 Code</a> •
    <a href="https://huggingface.co/datasets/WaitHZ/SSPO-data">🤗 Dataset</a>
  </p>
</div>


## 🌟Overview

Deep search agents operate over trajectories spanning dozens of information-seeking steps, but standard reinforcement learning provides only a single outcome reward for the entire trajectory. This sparse signal makes it difficult to determine which intermediate reasoning and tool-use actions should be reinforced or suppressed. Although on-policy self-distillation can provide denser supervision, directly distilling from a teacher with privileged information creates a severe information asymmetry: the teacher can take shortcuts that are unavailable to the student at inference time, leading to premature termination and tool-use collapse.

SSPO addresses this problem through two designs:

- **Evidence Anchors** are concise, step-level evidence snippets extracted from the web. They provide the self-teacher with action-relevant privileged information without prescribing a complete search trajectory.
- **Step-Level Self-Distilled Advantage Weights** convert teacher–student disagreement into a weight for each complete information-seeking step. The outcome reward determines the direction of the policy update, while the privileged teacher controls its magnitude. These weights are applied only to incorrect trajectories, leaving correct trajectories unaffected by the self-distillation signal.

Experiments with Qwen3-8B on BrowseComp, GAIA, and FRAMES show that SSPO consistently outperforms GRPO and matches or surpasses GRPO trained with twice as many gradient steps, while adding only about 5% training overhead from teacher scoring.

<div align="center">
  <img src="./overview.png" width="100%">
</div>

## 📚Dataset

The training data are available at [WaitHZ/SSPO-data](https://huggingface.co/datasets/WaitHZ/SSPO-data).

| File | Description | Used by |
| --- | --- | --- |
| `gpt-oss-120b-traj.jsonl` | Correct GPT-OSS-120B teacher trajectories | Cold-start SFT |
| `evidence_anchors.jsonl` | Questions, answers, and Evidence Anchors | GRPO and SSPO |

Download the files into `data/`:

```bash
hf download WaitHZ/SSPO-data \
  --repo-type dataset \
  --local-dir data
```

## 🚩Citation

```bib

```

## 🌻Acknowledgement

We thank [WebExplorer](https://github.com/hkust-nlp/WebExplorer) for its data-construction framework.