Mathematical Reasoning on OlympiadBench (MV, EM, PK)
37.86EM ScoreEntropy reweighting
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| Entropy reweightingBackbone=Qwen3-8B-Base, Layer=Deep (Layer 31)2026.05 | 37.86 | — | — | |
| Explorer-only, successful seeds (diagnostic)Backbone=Qwen3-8B-Base2026.05 | 37.18 | — | — | |
| Entropy reweightingBackbone=Qwen3-8B-Base, Layer=Shallow (Layer 8)2026.05 | 36.95 | — | — | |
| Entropy reweightingBackbone=Qwen3-8B-Base, Layer=Mid (Layer 20)2026.05 | 36.48 | — | — | |
| Full-token DAPOBackbone=Qwen3-8B-Base2026.05 | 34.24 | — | — | |
| High Prediction Entropy-onlyBackbone=Qwen3-8B-Base2026.05 | 33.67 | — | — | |
| Random-20%Backbone=Qwen3-8B-Base2026.05 | 32.83 | — | — | |
| Anchor-onlyBackbone=Qwen3-8B-Base2026.05 | 30.47 | — | — | |
| I-PPOBackbone=Phi3 4B2026.04 | 24.23 | 24.29 | 43.47 | |
| I-PPOBackbone=Qwen2.5 3B2026.04 | 22.11 | 23.28 | 44.03 | |
| I-PPOBackbone=LLaMa2026.04 | 21.59 | 22.87 | 43.1 | |
| I-PPOBackbone=LLaMa3 8B2026.04 | 21.59 | 22.87 | 43.1 | |
| Explorer-only, all seedsBackbone=Qwen3-8B-Base2026.05 | 20.18 | — | — | |
| I-PPOBackbone=Gemma2026.04 | 17.82 | 22.51 | 34.15 | |
| I-PPOBackbone=Gemma2 2B2026.04 | 17.82 | 22.51 | 34.15 | |
| PPOBackbone=Phi3 4B2026.04 | 16.63 | 21.33 | 41.35 | |
| PPOBackbone=LLaMa2026.04 | 16.49 | 20.94 | 41.19 | |
| PPOBackbone=LLaMa3 8B2026.04 | 16.49 | 20.94 | 41.19 | |
| GRPOBackbone=LLaMa2026.04 | 15.7 | 18.19 | 38.22 | |
| PPOBackbone=Qwen2.5 3B2026.04 | 15.32 | 20.61 | 41.03 | |
| PPOBackbone=Gemma2026.04 | 11.85 | 16.28 | 32.04 | |
| PPOBackbone=Gemma2 2B2026.04 | 11.85 | 16.28 | 32.04 | |
| BaselineBackbone=Phi3 4B2026.04 | 11.76 | 15.32 | 35.53 | |
| BaselineBackbone=LLaMa2026.04 | 10.58 | 15.72 | 35.81 | |
| BaselineBackbone=LLaMa3 8B2026.04 | 10.58 | 15.72 | 35.81 | |
| GRPOBackbone=Gemma2026.04 | 10.26 | 15.04 | 30.1 | |
| BaselineBackbone=Qwen2.5 3B2026.04 | 9.78 | 14.64 | 36.08 | |
| I-PPOBackbone=Rho2026.04 | 7.27 | 5.82 | 20.47 | |
| I-PPOBackbone=Rho 1B2026.04 | 7.27 | 5.82 | 20.47 | |
| BaselineBackbone=Gemma2026.04 | 6.4 | 10.55 | 28.52 | |
| BaselineBackbone=Gemma2 2B2026.04 | 6.4 | 10.55 | 28.52 | |
| PPOBackbone=Rho2026.04 | 4.45 | 5.76 | 20.21 | |
| PPOBackbone=Rho 1B2026.04 | 4.45 | 5.76 | 20.21 | |
| GRPOBackbone=Rho2026.04 | 3.59 | 3.12 | 15.71 | |
| BaselineBackbone=Rho2026.04 | 2.04 | 2.82 | 12.59 | |
| BaselineBackbone=Rho 1B2026.04 | 2.04 | 2.82 | 12.59 |