Code Generation on HumanEval 1024 sequence length (test)
48.78Code Generation ScoreAnchor-based History-stable Decoding (AHD)
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| Anchor-based History-stable Decoding (AHD)Sequence Length=1024, Backbone=LLaDA-8B-Instruct2026.04 | 48.78 | 6.99 | 146.48 | |
| LLaDASequence Length=1024, Backbone=LLaDA-8B-Instruct2026.04 | 43.9 | 1 | 1,024 | |
| KLASSSequence Length=1024, Backbone=LLaDA-8B-Instruct2026.04 | 43.9 | 3.3 | 309.97 | |
| SaberSequence Length=1024, Backbone=LLaDA-8B-Instruct2026.04 | 43.9 | 4.87 | 210.43 | |
| Fast-dLLMSequence Length=1024, Backbone=LLaDA-8B-Instruct2026.04 | 42.68 | 5.54 | 184.79 | |
| PC-samplerSequence Length=1024, Backbone=LLaDA-8B-Instruct2026.04 | 42.07 | 1 | 1,024 |