Code Generation on HumanEval (Accuracy and TPS)
61.7AccuracyFast-dLLM-v2
Evaluation results
| Method | Links | ||
|---|---|---|---|
| Fast-dLLM-v2Instruction Tuning=Dream-Instruct2026.04 | 61.7 | 2.58 | |
| TRIMSInstruction Tuning=Dream-Instruct, Sampling strategy=confidence-based sampling2026.04 | 57.3 | 2.21 | |
| d3LLMInstruction Tuning=Dream-Instruct2026.04 | 57.1 | 3.2 | |
| DreamInstruction Tuning=Dream-Instruct2026.04 | 55.2 | 1 | |
| Fast-dLLMInstruction Tuning=Dream-Instruct2026.04 | 54.3 | 1.33 | |
| dParallelInstruction Tuning=Dream-Instruct2026.04 | 54.3 | 2.57 |