# OPSD Experiment Results

Reproduction of [OPSD (On-Policy Self-Distillation)](https://github.com/siyan-zhao/OPSD) on Qwen3-1.7B, 4B, and 8B.

## Results (Avg@12)

### Qwen3-1.7B
| Method | AIME24 | AIME25 | HMMT25 |
|--------|:------:|:------:|:------:|
| Base | 47.2% | 35.3% | 21.9% |
| OPSD (best) | **49.2%** | **37.5%** | **24.4%** |
| SFT (best) | 37.5% | 30.8% | 19.2% |
| GRPO (best) | 47.8% | 35.0% | 22.8% |

### Qwen3-4B
| Method | AIME24 | AIME25 | HMMT25 |
|--------|:------:|:------:|:------:|
| Base | **71.1%** | 60.0% | 38.6% |
| OPSD (best) | 62.2% | 57.2% | 34.2% |
| SFT (best) | 62.5% | 58.1% | 33.3% |
| GRPO (best) | 68.9% | **65.0%** | **41.9%** |

### Qwen3-8B
| Method | AIME24 | AIME25 | HMMT25 |
|--------|:------:|:------:|:------:|
| Base | **72.8%** | 61.7% | 38.6% |
| OPSD (best) | 69.4% | 63.3% | 38.6% |
| SFT (best) | 69.2% | 60.3% | 36.1% |
| GRPO (best) | 72.2% | **65.8%** | **40.8%** |

## Setup
- All methods: lr=5e-6, BS=32, LoRA r=64 alpha=128, 200 steps
- Eval: val_n=12, temperature=1.0, thinking mode enabled
- Data: siyanzhao/Openthoughts_math_30k_opsd

## Reference
[Self-Distilled Reasoner: On-Policy Self-Distillation for LLMs](https://arxiv.org/pdf/2601.18734v3)