Code Generation on Energy-Aware Code Generation (Held-out problems)
63.5Train Success RateEnergy-SFT + EDP
Evaluation Results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| Energy-SFT + EDPInitialization=E-SFT, Reward Function=EDP (Energy-Delay Product), Configuration Group=Multi-objective2026.07 | 63.5 | 81.7 | 44.6 | 34.2 | 58.4 | 4.45 | 12.63 | |
| Energy-SFT + energyInitialization=E-SFT, Reward Function=energy, Configuration Group=Pure single-axis vs. baseline2026.07 | 53.4 | 82 | 51.1 | 30.71 | 33.4 | 4.45 | 12.78 | |
| Energy-SFT + runtimeInitialization=E-SFT, Reward Function=runtime, Configuration Group=Cross-objective exploration2026.07 | 53.2 | 82.2 | 51.1 | 30.54 | 33.3 | 4.45 | 14.44 | |
| AfterburnerInitialization=Runtime-SFT, Reward Function=runtime, Configuration Group=Pure single-axis vs. baseline2026.07 | 49.3 | 88.5 | 64.6 | 23.37 | 22.1 | 7.64 | 12.19 | |
| Runtime-SFT + EDPInitialization=Runtime-SFT, Reward Function=EDP (Energy-Delay Product), Configuration Group=Multi-objective2026.07 | 34.6 | 86.3 | 40.8 | 46.93 | 51.5 | 7.64 | 16.41 |