Science and Question Answering on ScienceQA, SampleQA, GPQA, HLE
1.91SampleQA ScorePPO
Evaluation results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| PPOBackbone=Qwen3-1.7B, Supervision Type=Noisy supervision2026.02 | 1.91 | 2.54 | 3.94 | 2.8 | 32.45 | |
| BaseBackbone=Qwen3-1.7B, Supervision Type=Noisy supervision2026.02 | 1.78 | 2.36 | 2.87 | 2.55 | 32.13 | |
| FlowRLBackbone=Qwen3-1.7B, Supervision Type=Noisy supervision2026.02 | 1.76 | 1.81 | 3.99 | 2.52 | 32.3 | |
| DFPOBackbone=Qwen3-1.7B, Supervision Type=Noisy supervision2026.02 | 1.76 | 2.54 | 3.99 | 2.76 | 33.75 | |
| KTAEBackbone=Qwen3-1.7B, Supervision Type=Noisy supervision2026.02 | 1.64 | 2.9 | 4.36 | 2.97 | 29.41 | |
| Robust BellmanBackbone=Qwen3-1.7B, Supervision Type=Noisy supervision2026.02 | 1.64 | 2.17 | 4.17 | 2.66 | 31.72 | |
| λ-GRPOBackbone=Qwen3-1.7B, Supervision Type=Noisy supervision2026.02 | 1.62 | 2.9 | 3.89 | 2.8 | 32.33 | |
| GRPOBackbone=Qwen3-1.7B, Supervision Type=Noisy supervision2026.02 | 1.6 | 1.99 | 3.94 | 2.51 | 31.7 | |
| BAPOBackbone=Qwen3-1.7B, Supervision Type=Noisy supervision2026.02 | 1.55 | 2.36 | 3.48 | 2.46 | 29.12 | |
| Dr. GRPOBackbone=Qwen3-1.7B, Supervision Type=Noisy supervision2026.02 | 1.5 | 2.17 | 3.1 | 2.26 | 32.18 | |
| Reinforce++Backbone=Qwen3-1.7B, Supervision Type=Noisy supervision2026.02 | 1.48 | 2.36 | 3.52 | 2.45 | 32.75 |