aiqbench analysis
This research addresses a critical bottleneck in diffusion language models by proposing a training-free method to construct "soft tokens" that enforce coherence during parallel decoding. By correcting a geometric mismatch in the embedding space, the authors demonstrate that uncertainty preservation alone is insufficient and that favoring consistent token sequences significantly improves generation quality on math and code benchmarks. The findings offer immediate practical value for deploying efficient parallel inference without requiring additional model training.
Rethinking Soft Tokens for Parallel Decoding in Diffusion Language Models
arXiv:2609.37391v1 Announce Type: new Abstract: Diffusion language models (DLMs) enable parallel generation by predicting and committing multiple tokens at each denoising step, yet they can generate individually plausible but mutually inconsistent tokens. Recent work shows that \emph{soft tokens} can mitigate this issue by representing uncertain positions with continuous embeddings built from the model's predictive distribution at the previous decoding step. However, although soft tokens are commonly understood as preserving predictive uncertainty, how soft-token feedback improves parallel decoding has not been systematically examined. In this paper, we investigate this question in frozen pretrained DLMs to examine soft-token feedback without the effects of additional training. To construct soft-token inputs in a training-free setting, we identify a geometric mismatch between conventional soft-token construction and the pretrained embedding space. Based on this observation, we propose a training-free, geometry-aware construction of soft tokens. Our analysis of soft-token feedback suggests that uncertainty preservation alone does not fully explain how it reshapes subsequent predictions. To better explain how soft-token feedback improves parallel decoding, we provide empirical evidence that it favors coherent token sequences. Across four pretrained DLMs and four math and code benchmarks, our method outperforms standard parallel decoding and a training-free Euclidean soft-token baseline. Code: https://github.com/kodaikawamura/rethinking-soft-tokens
Abstract:Diffusion language models (DLMs) enable parallel generation by predicting and committing multiple tokens at each denoising step, yet they can generate individually plausible but mutually inconsistent tokens. Recent work shows that \emph{soft tokens} can mitigate this issue by representing uncertain positions with continuous embeddings built from the model's predictive distribution at the previous decoding step. However, although soft tokens are commonly understood as preserving predictive uncertainty, how soft-token feedback improves parallel decoding has not been systematically examined. In this paper, we investigate this question in frozen pretrained DLMs to examine soft-token feedback without the effects of additional training. To construct soft-token inputs in a training-free setting, we identify a geometric mismatch between conventional soft-token construction and the pretrained embedding space. Based on this observation, we propose a training-free, geometry-aware construction of soft tokens. Our analysis of soft-token feedback suggests that uncertainty preservation alone does not fully explain how it reshapes subsequent predictions. To better explain how soft-token feedback improves parallel decoding, we provide empirical evidence that it favors coherent token sequences. Across four pretrained DLMs and four math and code benchmarks, our method outperforms standard parallel decoding and a training-free Euclidean soft-token baseline. Code: this https URL
Machine Learning (cs.LG)
arXiv:2609.37391 [cs.LG]
(or arXiv:2609.37391v1 [cs.LG] for this version)
https://doi.org/10.48550/arXiv.2609.37391 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 291 words
Related coverage
Attention-Discounted Adaptive Sampler for Masked Diffusion Language Models
arXiv · Jun 10, 2026
Rethinking the Generation Order of Block Diffusion Language Models
arXiv · Jul 28, 2026
Rethinking the Roles of Diffusion and Decoding: Predictive-State Flows for Parallel Language Generation
arXiv · Oct 12, 2026
Academia is for Ambition — Alex Zhang, MIT
Latent Space · Oct 2, 2026