Reward Modeling on Skywork-Reward-Preference (test)
93.6Final Token AccuracyTCRM
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| TCRMBase Model=Qwen3 14B, Reward Model Type=TCRM (ours)2026.04 | 93.6 | 88.7 | 0.02 | 0.85 | |
| TCRMBase Model=Qwen3 32B, Reward Model Type=TCRM (ours)2026.04 | 93.6 | 88.9 | 0.01 | 0.34 | |
| BaselineBase Model=Qwen3 32B, Reward Model Type=Baseline2026.04 | 93.5 | 60.1 | 0.99 | 2.1 | |
| BaselineBase Model=Qwen3 14B, Reward Model Type=Baseline2026.04 | 93.2 | 60.1 | 1.79 | 5.16 | |
| BaselineBase Model=Llama 3.1 8B, Reward Model Type=Baseline2026.04 | 92.6 | 51.1 | 9.18 | 13.62 | |
| TC-λBase Model=Qwen3 14B, Reward Model Type=TC-λ2026.04 | 92.5 | 86.3 | 0.04 | 1.05 | |
| TCRMBase Model=Llama 3.1 8B, Reward Model Type=TCRM (ours)2026.04 | 92.5 | 84.7 | 0.06 | 1.22 | |
| TC-λBase Model=Llama 3.1 8B, Reward Model Type=TC-λ2026.04 | 92.3 | 71.7 | 0.08 | 2.88 | |
| TCRMBase Model=Qwen3 8B, Reward Model Type=TCRM (ours)2026.04 | 90.4 | 74.3 | 0.23 | 2.22 | |
| BaselineBase Model=Qwen3 4B, Reward Model Type=Baseline2026.04 | 90.3 | 50.4 | 8.12 | 20.58 | |
| BaselineBase Model=Qwen3 8B, Reward Model Type=Baseline2026.04 | 90.3 | 52 | 4.27 | 10.68 | |
| TC-λBase Model=Qwen3 8B, Reward Model Type=TC-λ2026.04 | 90 | 62 | 0.34 | 3.88 | |
| TC-λBase Model=Qwen3 4B, Reward Model Type=TC-λ2026.04 | 89.7 | 64.5 | 0.36 | 3.25 | |
| TCRMBase Model=Qwen3 4B, Reward Model Type=TCRM (ours)2026.04 | 89.7 | 78.1 | 0.26 | 1.68 | |
| BaselineBase Model=Qwen3 1.7B, Reward Model Type=Baseline2026.04 | 86.5 | 52.3 | 3.91 | 6.22 | |
| TC-λBase Model=Qwen3 1.7B, Reward Model Type=TC-λ2026.04 | 86.5 | 54.9 | 0.64 | 3.65 | |
| TCRMBase Model=Qwen3 1.7B, Reward Model Type=TCRM (ours)2026.04 | 86.5 | 57.6 | 0.67 | 3.19 | |
| ImplicitPRM (DPO)Base Model=Llama 3.1 8B, Reward Model Type=ImplicitPRM (DPO)2026.04 | 84.4 | 79.8 | 0.02 | 56.72 | |
| TCRMBase Model=Qwen3 0.6B, Reward Model Type=TCRM (ours)2026.04 | 84.2 | 63.5 | 0.64 | 2.18 | |
| TC-λBase Model=Qwen3 0.6B, Reward Model Type=TC-λ2026.04 | 84 | 52.6 | 0.62 | 3.24 | |
| BaselineBase Model=Qwen3 0.6B, Reward Model Type=Baseline2026.04 | 83.9 | 48 | 5.35 | 10.7 | |
| ImplicitPRM (DPO)Base Model=Qwen3 14B, Reward Model Type=ImplicitPRM (DPO)2026.04 | 83.4 | 79.7 | 0 | 4.88 | |
| ImplicitPRM (DPO)Base Model=Qwen3 4B, Reward Model Type=ImplicitPRM (DPO)2026.04 | 83.3 | 78 | 0 | 14.67 | |
| ImplicitPRM (DPO)Base Model=Qwen3 8B, Reward Model Type=ImplicitPRM (DPO)2026.04 | 83 | 77.8 | 0 | 8.73 | |
| ImplicitPRM (DPO)Base Model=Qwen3 32B, Reward Model Type=ImplicitPRM (DPO)2026.04 | 82.6 | 78.3 | 0 | 4.78 | |
| ImplicitPRM (DPO)Base Model=Qwen3 1.7B, Reward Model Type=ImplicitPRM (DPO)2026.04 | 79.8 | 75.3 | 0 | 6.98 | |
| ImplicitPRM (DPO)Base Model=Qwen3 0.6B, Reward Model Type=ImplicitPRM (DPO)2026.04 | 78.2 | 73.6 | 0 | 5.59 |