General Reasoning on Out-of-Distribution Multi-domain Reasoning Winogrande GPQA MuSR
82.8Winogrande AccuracyTP-Schedule (8B)
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| TP-Schedule (8B)Backbone=Qwen3-8B, Credit Allocation Method=GRPO, Temporal Scheduling=true, Trajectory Percentile (TP)=true2026.05 | 82.8 | 64.1 | 64.5 | 70.5 | |
| GRPO (8B)Backbone=Qwen3-8B, Credit Allocation Method=GRPO2026.05 | 82 | 62.7 | 63.9 | 69.5 | |
| Qwen3-8BBackbone=Qwen3-8B2026.05 | 81.3 | 59.6 | 62.6 | 67.9 | |
| Entropy Adv. + ScheduleBackbone=Qwen3-4B, Credit Allocation Method=Entropy Adv., Temporal Scheduling=true2026.05 | 78.5 | 55.6 | 62.2 | 65.4 | |
| TP-ScheduleBackbone=Qwen3-4B, Credit Allocation Method=GRPO, Temporal Scheduling=true, Trajectory Percentile (TP)=true2026.05 | 78.4 | 60.1 | 62.9 | 67.1 | |
| Forking Tok. + ScheduleBackbone=Qwen3-4B, Credit Allocation Method=Forking Tok., Temporal Scheduling=true2026.05 | 78.2 | 56.3 | 63.2 | 65.9 | |
| TP-Schedule (GSPO)Backbone=Qwen3-4B, Credit Allocation Method=GSPO, Temporal Scheduling=true, Trajectory Percentile (TP)=true2026.05 | 78.1 | 57.1 | 62.4 | 65.9 | |
| Entropy Adv.Backbone=Qwen3-4B, Credit Allocation Method=Entropy Adv.2026.05 | 77.8 | 54.5 | 61.7 | 64.6 | |
| Qwen3-4BBackbone=Qwen3-4B2026.05 | 77.7 | 54.1 | 61.4 | 64.4 | |
| Forking Tok.Backbone=Qwen3-4B, Credit Allocation Method=Forking Tok.2026.05 | 77.6 | 55.6 | 61.5 | 64.9 | |
| GRPOBackbone=Qwen3-4B, Credit Allocation Method=GRPO2026.05 | 76.8 | 55.6 | 60.9 | 64.4 | |
| GSPOBackbone=Qwen3-4B, Credit Allocation Method=GSPO2026.05 | 76.6 | 54.5 | 61.5 | 64.2 |