Unified Energy for Invariant and Independent Decoding in Diffusion Language Models
About
Diffusion Language Models (DLMs) enable parallel text generation by iteratively denoising a full sequence, offering attractive flexibility compared to auto-regressive (AR) decoding. However, existing methods fail to fully capture token relationships, leading to a performance gap relative to AR baselines, especially as the degree of parallelism increases. In this paper, we give a systematic analysis of the gap, identifying three key factors: (i) model capacity, (ii) dependency, and (iii) invariance. To address these issues, we first propose an invariant energy (Inv-E) together with an effective sampling-based estimator to handle the invariance issue. By further combining with the independent energy (Ind-E), we obtain a unified energy (Uni-E), that accounts for all these factors. Uni-E enjoys a unique advantage: it can be computed exactly without sampling-based partition estimation. Besides, Uni-E is model agnostic and can therefore be scaled to models of arbitrary size. We further prove that Uni-E can correct the distribution shift caused by dependency and invariance. Extensive experiments across Diffusion Language Models (DLMs) and Diffusion Large Language Models (DLLMs) demonstrate the effectiveness of the proposed Uni-E.
Related benchmarks
| Task | Dataset | Result | Rank | |
|---|---|---|---|---|
| Commonsense Reasoning | HellaSwag | HellaSwag Accuracy73.7 | 897 | |
| Mathematical Reasoning | MATH 500 | Accuracy39.2 | 589 | |
| Multitask Language Understanding | MMLU | Accuracy66.5 | 568 | |
| Language Modeling | OpenWebText | Perplexity16.54 | 190 | |
| Multi-task Language Understanding | MMLU | Top-1 Accuracy66.5 | 46 | |
| Language Modeling | LAMBADA zero-shot (test) | -- | 44 | |
| Language Generation | Language Modeling Evaluation Set | Generative Perplexity (Llama2)13.1 | 42 | |
| Reasoning | LiveBench | Accuracy10.9 | 40 | |
| Language Modeling | PTB zero-shot | Perplexity86.18 | 35 | |
| Language Modeling | Pubmed zero-shot | Perplexity41.35 | 30 |