LLM Inference Efficiency Optimization on Held-out (test)
0.024TFLOPsOracle
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| OracleBackbone=LLaDA-8B, Condition=L = k2026.05 | 0.024 | 99.4 | 0 | |
| Predict-then-DiffuseBackbone=LLaDA-8B2026.05 | 0.026 | 99.34 | 0.1 | |
| Doubling HeuristicBackbone=LLaDA-8B, Strategy=Mean Doubling2026.05 | 0.027 | 99.31 | — | |
| Static Response LengthBackbone=LLaDA-8B, Length Limit=2002026.05 | 0.054 | 98.6 | 22.1 | |
| Max Response LengthBackbone=LLaDA-8B, Length Limit=40962026.05 | 4.03 | 0 | 0 |