Code Generation on LCB
7.81SpeedupBASTION
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| BASTIONModel=Qwen3-4B, Decoding Mode=Greedy Decoding, Temperature=02026.05 | 7.81 | — | 9.48 | — | |
| BASTIONModel=Qwen3-8B, Decoding Mode=Stochastic Decoding, Temperature=12026.05 | 7.81 | — | 9.82 | — | |
| BASTIONModel=Qwen3-4B, Decoding Mode=Stochastic Decoding, Temperature=12026.05 | 7.69 | — | 9.83 | — | |
| JetFlowBudget=256, Temperature=02026.06 | 7.67 | — | 8.79 | — | |
| BASTIONModel=Qwen3-8B, Decoding Mode=Greedy Decoding, Temperature=02026.05 | 7.64 | — | 9.91 | — | |
| JetFlowBudget=128, Temperature=02026.06 | 7.29 | — | 8.21 | — | |
| JetFlowBudget=256, Temperature=12026.06 | 7.25 | — | 8.29 | — | |
| JetFlowBudget=128, Temperature=12026.06 | 6.84 | — | 7.74 | — | |
| DFlash-TBudget=256, Temperature=02026.06 | 6.75 | — | 7.72 | — | |
| DFlash-TBudget=128, Temperature=02026.06 | 6.42 | — | 7.25 | — | |
| DFlash-TBudget=256, Temperature=12026.06 | 6.26 | — | 7.17 | — | |
| DFlash-TBudget=128, Temperature=12026.06 | 5.99 | — | 6.71 | — | |
| JetFlowBudget=64, Temperature=02026.06 | 5.95 | — | 6.59 | — | |
| DFlashModel=Qwen3-8B, Decoding Mode=Greedy Decoding, Temperature=02026.05 | 5.9 | — | 7.64 | — | |
| JetFlowBudget=64, Temperature=12026.06 | 5.76 | — | 6.38 | — | |
| DFlashModel=Qwen3-4B, Decoding Mode=Greedy Decoding, Temperature=02026.05 | 5.71 | — | 7.23 | — | |
| DFlashTemperature=0, Model=Qwen3-4B, Draft size=162026.05 | 5.48 | — | 7.15 | — | |
| DFlash-TBudget=64, Temperature=02026.06 | 5.47 | — | 6.06 | — | |
| DFlashTemperature=0, Model=Qwen3-8B, Draft size=162026.05 | 5.37 | — | 7.12 | — | |
| DOMINOTemperature=0, Model=Qwen3-4B, Draft size=162026.05 | 5.33 | — | 6.86 | — | |
| DFlash-TBudget=64, Temperature=12026.06 | 5.28 | — | 5.82 | — | |
| DOMINOTemperature=0, Model=Qwen3-8B, Draft size=162026.05 | 5.27 | — | 7.04 | — | |
| DFlashModel=Qwen3-8B, Decoding Mode=Stochastic Decoding, Temperature=12026.05 | 5.21 | — | 6.91 | — | |
| DFlashModel=Qwen3-4B, Decoding Mode=Stochastic Decoding, Temperature=12026.05 | 5.1 | — | 6.9 | — | |
| DOMINOTemperature=1, Model=Qwen3-8B, Draft size=162026.05 | 5.06 | — | 6.72 | — | |
| DOMINOTemperature=1, Model=Qwen3-4B, Draft size=162026.05 | 5.04 | — | 6.45 | — | |
| DFlashTemperature=1, Model=Qwen3-8B, Draft size=162026.05 | 5.04 | — | 6.73 | — | |
| JetFlowBudget=32, Temperature=02026.06 | 4.86 | — | 6.48 | — | |
| JetFlowBudget=16, Temperature=02026.06 | 4.77 | — | 6.24 | — | |
| DFlashTemperature=1, Model=Qwen3-4B, Draft size=162026.05 | 4.76 | — | 6.38 | — | |
| DFlashBudget=16, Temperature=02026.06 | 4.7 | — | 6.16 | — | |
| JetFlowBudget=32, Temperature=12026.06 | 4.53 | — | 5.93 | — | |
| JetFlowBudget=16, Temperature=12026.06 | 4.41 | — | 5.74 | — | |
| DFlashBudget=16, Temperature=12026.06 | 4.38 | — | 5.7 | — | |
| DFlashBudget=32, Temperature=02026.06 | 3.99 | — | 5.23 | — | |
| DFlashBudget=32, Temperature=12026.06 | 3.74 | — | 4.81 | — | |
| EAGLE-3Model=Qwen3-4B, Decoding Mode=Greedy Decoding, Temperature=02026.05 | 2.66 | — | 2.1 | — | |
| DFlashModel=Qwen3-Coder-30B-A3B, Concurrency=1, Inference Engine=SGLang, Backend=FA42026.02 | 2.6 | 569 | 6.42 | — | |
| EAGLE-3Model=Qwen3-4B, Decoding Mode=Stochastic Decoding, Temperature=12026.05 | 2.38 | — | 2.08 | — | |
| HyperDFlashTemperature=0, Number of drafted steps=6, Mode=Non-thinking2026.06 | 2.38 | — | 3.4 | — | |
| EAGLE-3Model=Qwen3-8B, Decoding Mode=Greedy Decoding, Temperature=02026.05 | 2.29 | — | 1.91 | — | |
| EAGLE-3Model=Qwen3-8B, Decoding Mode=Stochastic Decoding, Temperature=12026.05 | 2.28 | — | 1.84 | — | |
| DARTTemperature=1, Model=Qwen3-8B, Draft size=602026.05 | 2.25 | — | 2.78 | — | |
| DARTTemperature=0, Model=Qwen3-8B, Draft size=602026.05 | 2.24 | — | 2.78 | — | |
| DARTTemperature=0, Model=Qwen3-4B, Draft size=602026.05 | 2.23 | — | 2.76 | — | |
| DARTTemperature=1, Model=Qwen3-4B, Draft size=602026.05 | 2.23 | — | 2.76 | — | |
| EAGLE-3Temperature=0, Model=Qwen3-4B, Draft size=602026.05 | 2.19 | — | 3.31 | — | |
| HyperDFlashTemperature=0, Drafted steps per verification round=6, Decoding mode=Think-high mode2026.06 | 2.18 | — | 2.88 | — | |
| HyperDFlashTemperature=1, Number of drafted steps=6, Mode=Non-thinking2026.06 | 2.17 | — | 3.16 | — | |
| EAGLE-3Budget=64, Temperature=02026.06 | 2.09 | — | 3.62 | — | |
| MTPTemperature=0, Drafted steps per verification round=3, Decoding mode=Think-high mode2026.06 | 2.09 | — | 2.7 | — | |
| MTPTemperature=0, Number of drafted steps=3, Mode=Non-thinking2026.06 | 2.06 | — | 2.87 | — | |
| EAGLE-3Temperature=1, Model=Qwen3-4B, Draft size=602026.05 | 2.05 | — | 3.19 | — | |
| EAGLE-3Temperature=0, Model=Qwen3-8B, Draft size=602026.05 | 2.03 | — | 3.12 | — | |
| MTPTemperature=1, Number of drafted steps=3, Mode=Non-thinking2026.06 | 2.02 | — | 2.73 | — | |
| EAGLE-3Budget=64, Temperature=12026.06 | 1.99 | — | 3.49 | — | |
| EAGLE-3Budget=32, Temperature=02026.06 | 1.96 | — | 3.42 | — | |
| EAGLE-3Temperature=0, Model=Qwen3-4B, Draft size=162026.05 | 1.95 | — | 2.93 | — | |
| HyperDFlashTemperature=1, Drafted steps per verification round=6, Decoding mode=Think-high mode2026.06 | 1.95 | — | 2.55 | — | |
| EAGLE-3Budget=32, Temperature=12026.06 | 1.91 | — | 3.31 | — | |
| EAGLE-3Temperature=1, Model=Qwen3-8B, Draft size=162026.05 | 1.9 | — | 2.93 | — | |
| EAGLE-3Temperature=1, Model=Qwen3-8B, Draft size=602026.05 | 1.9 | — | 2.95 | — | |
| MTPTemperature=1, Drafted steps per verification round=3, Decoding mode=Think-high mode2026.06 | 1.89 | — | 2.44 | — | |
| EAGLE-3Temperature=1, Model=Qwen3-4B, Draft size=162026.05 | 1.87 | — | 2.83 | — | |
| EAGLE-3Budget=16, Temperature=02026.06 | 1.82 | — | 3.19 | — | |
| EAGLE-3Budget=16, Temperature=12026.06 | 1.81 | — | 3.09 | — | |
| EAGLE-3Temperature=0, Model=Qwen3-8B, Draft size=162026.05 | 1.8 | — | 2.95 | — | |
| MTPTemperature=0, Number of drafted steps=6, Mode=Non-thinking2026.06 | 1.6 | — | 3.01 | — | |
| MTPTemperature=0, Drafted steps per verification round=6, Decoding mode=Think-high mode2026.06 | 1.56 | — | 2.79 | — | |
| MTPTemperature=1, Number of drafted steps=6, Mode=Non-thinking2026.06 | 1.5 | — | 2.79 | — | |
| Vanilla DFlashTemperature=0, Number of drafted steps=6, Mode=Non-thinking2026.06 | 1.48 | — | 1.93 | — | |
| Vanilla DFlashTemperature=1, Number of drafted steps=6, Mode=Non-thinking2026.06 | 1.42 | — | 1.87 | — | |
| Vanilla DFlashTemperature=0, Drafted steps per verification round=6, Decoding mode=Think-high mode2026.06 | 1.4 | — | 1.8 | — | |
| MTPTemperature=1, Drafted steps per verification round=6, Decoding mode=Think-high mode2026.06 | 1.4 | — | 2.51 | — | |
| Vanilla DFlashTemperature=1, Drafted steps per verification round=6, Decoding mode=Think-high mode2026.06 | 1.34 | — | 1.71 | — | |
| BaselineModel=Qwen3-Coder-30B-A3B, Concurrency=1, Inference Engine=SGLang, Backend=FA42026.02 | — | 220 | — | — | |
| DFlashtarget_model=Qwen3-4B2026.07 | — | — | — | 4.18 | |
| DFlashtarget_model=Qwen3-8B2026.07 | — | — | — | 4.39 | |
| DFlashtarget_model=Qwen3-14B2026.07 | — | — | — | 4.33 | |
| DFlashtarget_model=Gemma4-12B2026.07 | — | — | — | 3.7 | |
| DSparktarget_model=Qwen3-4B2026.07 | — | — | — | 4.86 | |
| DSparktarget_model=Qwen3-8B2026.07 | — | — | — | 5.17 | |
| DSparktarget_model=Qwen3-14B2026.07 | — | — | — | 5.02 | |
| DSparktarget_model=Gemma4-12B2026.07 | — | — | — | 4.51 | |
| Eagle3target_model=Qwen3-4B2026.07 | — | — | — | 3.77 | |
| Eagle3target_model=Qwen3-8B2026.07 | — | — | — | 4.17 | |
| Eagle3target_model=Qwen3-14B2026.07 | — | — | — | 4.01 | |
| Eagle3target_model=Gemma4-12B2026.07 | — | — | — | 4.16 |