Code Generation on LiveCodeBench (Pass@1, Step, Time)
11Pass@1Saber
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| SaberSampling Strategy=Efficient DLM Sampling, Backbone=LLaDA-8B-Instruct2025.10 | 11 | 122.47 | 2 | |
| ConfidenceSampling Strategy=Standard DLM Sampling, Backbone=LLaDA-8B-Instruct2025.10 | 9.75 | 256 | 5 | |
| SAR (p=2)Sampling Strategy=Efficient DLM Sampling, p parameter=2, Backbone=LLaDA-8B-Instruct2025.10 | 9.5 | 128 | 2 | |
| Confidence (p=2)Sampling Strategy=Efficient DLM Sampling, p parameter=2, Backbone=LLaDA-8B-Instruct2025.10 | 9.25 | 128 | 2 | |
| WINOSampling Strategy=Efficient DLM Sampling, Backbone=LLaDA-8B-Instruct2025.10 | 9.25 | 77.43 | 2 | |
| Fast-dLLMSampling Strategy=Efficient DLM Sampling, Backbone=LLaDA-8B-Instruct2025.10 | 8.75 | 256 | 2 | |
| EntropySampling Strategy=Standard DLM Sampling, Backbone=LLaDA-8B-Instruct2025.10 | 4 | 256 | 4 | |
| ReMDMSampling Strategy=Efficient DLM Sampling, Backbone=LLaDA-8B-Instruct2025.10 | 3.3 | 128 | 2 | |
| Fast-dLLM (+parallel)Sampling Strategy=Efficient DLM Sampling, Parallelism=true, Backbone=LLaDA-8B-Instruct2025.10 | 2.3 | 96.28 | 43 | |
| RandomSampling Strategy=Standard DLM Sampling, Backbone=LLaDA-8B-Instruct2025.10 | 0 | 256 | 4 |