Mathematical Reasoning on GSM8K (Speedup, Average Accepted Length (Tau))
7.82SpeedupJetFlow
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| JetFlowBudget=256, Temperature=0, Backbone Model=Qwen3-8B2026.06 | 7.82 | 8.62 | |
| JetFlowBudget=128, Temperature=0, Backbone Model=Qwen3-8B2026.06 | 7.34 | 8.05 | |
| JetFlowBudget=256, Temperature=1, Backbone Model=Qwen3-8B2026.06 | 7.16 | 8.03 | |
| DFlash-TBudget=256, Temperature=0, Backbone Model=Qwen3-8B2026.06 | 7.04 | 7.77 | |
| JetFlowBudget=128, Temperature=1, Backbone Model=Qwen3-8B2026.06 | 6.76 | 7.49 | |
| DFlash-TBudget=128, Temperature=0, Backbone Model=Qwen3-8B2026.06 | 6.63 | 7.31 | |
| DFlash-TBudget=256, Temperature=1, Backbone Model=Qwen3-8B2026.06 | 6.41 | 7.17 | |
| DFlash-TBudget=128, Temperature=1, Backbone Model=Qwen3-8B2026.06 | 6.11 | 6.72 | |
| JetFlowBudget=64, Temperature=0, Backbone Model=Qwen3-8B, Decoding Mode=causal tree drafting2026.06 | 5.98 | 6.56 | |
| DFlash-TBudget=64, Temperature=0, Backbone Model=Qwen3-8B, Decoding Mode=branch-agnostic tree construction2026.06 | 5.63 | 6.18 | |
| JetFlowBudget=64, Temperature=1, Backbone Model=Qwen3-8B2026.06 | 5.63 | 6.19 | |
| DFlash-TBudget=64, Temperature=1, Backbone Model=Qwen3-8B2026.06 | 5.28 | 5.77 | |
| EAGLE-3Budget=64, Temperature=0, Backbone Model=Qwen3-8B, Decoding Mode=tree mode (max depth 8)2026.06 | 2.53 | 4.31 | |
| EAGLE-3Budget=64, Temperature=1, Backbone Model=Qwen3-8B2026.06 | 2.35 | 4.17 |