Graduate QA on GPQA (held-out)
31.5Accuracy (GPQA held-out)Baseline
Evaluation Results
| Method | Links | |
|---|---|---|
| BaselineModel=Qwen3.5-2B2026.06 | 31.5 | |
| Random StringModel=Qwen3.5-2B2026.06 | 30.77 | |
| LoRA (DAPO)Model=Qwen3.5-2B2026.06 | 30.04 | |
| Random ImageModel=Qwen3.5-2B2026.06 | 27.47 | |
| Optimized Image (ART)Model=Qwen3.5-2B2026.06 | 26.37 | |
| LoRA (DAPO)Model=Qwen3.5-0.8B2026.06 | 24.18 | |
| Fixed Seed ImageModel=Qwen3.5-2B2026.06 | 24.18 | |
| BaselineModel=Qwen3.5-0.8B2026.06 | 23.44 | |
| Optimized Image (ART)Model=Qwen3.5-0.8B2026.06 | 20.15 | |
| Random ImageModel=Qwen3.5-0.8B2026.06 | 19.78 | |
| Random StringModel=Qwen3.5-0.8B2026.06 | 19.41 | |
| Fixed Seed ImageModel=Qwen3.5-0.8B2026.06 | 16.85 |