Open-ended Text Generation on MT-Bench
4.88MT-1 ScoreDECA
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| DECABackbone=Qwen2-1.5B2026.06 | 4.88 | — | 3.3 | 4.1 | |
| DeCAFBackbone=Qwen2-1.5B2026.06 | 4.87 | — | 3.26 | 4.06 | |
| DECABackbone=Llama-2-7B2026.06 | 4.5 | — | 2.93 | 3.72 | |
| Dec-LoRABackbone=Qwen2-1.5B2026.06 | 4.43 | — | 2.95 | 3.69 | |
| DeCAFBackbone=Llama-2-7B2026.06 | 4.31 | — | 3 | 3.65 | |
| Dec-LoRABackbone=Llama-2-7B2026.06 | 4.3 | — | 2.87 | 3.59 | |
| Dec-AdapterBackbone=Qwen2-1.5B2026.06 | 4.18 | — | 2.65 | 3.41 | |
| Dec-AdapterBackbone=Llama-2-7B2026.06 | 3.92 | — | 2.62 | 3.27 | |
| CE-only baselineTraining Objective=CE Only, Base Model=LLaDA-8B-Instruct2026.01 | — | 2.2 | — | — | |
| CTC-trained MDLMTraining Objective=CTC + CE, Base Model=LLaDA-8B-Instruct2026.01 | — | 3.7 | — | — | |
| LLaDA-1.5Base Model=LLaDA-1.52026.01 | — | 3.3 | — | — | |
| LLaDA-8B-InstructBase Model=LLaDA-8B-Instruct2026.01 | — | 2.8 | — | — |