Optimizing Decoding Paths in Masked Diffusion Models by Quantifying Uncertainty

Fuente: arXiv
Salvato in:
Dettagli Bibliografici
Autori principali: Chen, Ziyu, Jiang, Xinbei, Sun, Peng, Lin, Tao
Natura: Preprint
Pubblicazione: 2025
Soggetti:
Accesso online:
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866908731289108480
author Chen, Ziyu
Jiang, Xinbei
Sun, Peng
Lin, Tao
author_facet Chen, Ziyu
Jiang, Xinbei
Sun, Peng
Lin, Tao
contents Masked Diffusion Models (MDMs) offer flexible, non-autoregressive generation, but this freedom introduces a challenge: final output quality is highly sensitive to the decoding order. We are the first to formalize this issue, attributing the variability in output quality to the cumulative predictive uncertainty along a generative path. To quantify this uncertainty, we introduce Denoising Entropy, a computable metric that serves as an internal signal for evaluating generative process. Leveraging this metric, we propose two algorithms designed to optimize the decoding path: a post-hoc selection method and a real-time guidance strategy. Experiments demonstrate that our entropy-guided methods significantly improve generation quality, consistently boosting accuracy on challenging reasoning, planning, and code benchmarks. Our work establishes Denoising Entropy as a principled tool for understanding and controlling generation, effectively turning the uncertainty in MDMs from a liability into a key advantage for discovering high-quality solutions.
format Preprint
id arxiv_https___arxiv_org_abs_2512_21336
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Optimizing Decoding Paths in Masked Diffusion Models by Quantifying Uncertainty
Chen, Ziyu
Jiang, Xinbei
Sun, Peng
Lin, Tao
Computation and Language
Artificial Intelligence
Machine Learning
Masked Diffusion Models (MDMs) offer flexible, non-autoregressive generation, but this freedom introduces a challenge: final output quality is highly sensitive to the decoding order. We are the first to formalize this issue, attributing the variability in output quality to the cumulative predictive uncertainty along a generative path. To quantify this uncertainty, we introduce Denoising Entropy, a computable metric that serves as an internal signal for evaluating generative process. Leveraging this metric, we propose two algorithms designed to optimize the decoding path: a post-hoc selection method and a real-time guidance strategy. Experiments demonstrate that our entropy-guided methods significantly improve generation quality, consistently boosting accuracy on challenging reasoning, planning, and code benchmarks. Our work establishes Denoising Entropy as a principled tool for understanding and controlling generation, effectively turning the uncertainty in MDMs from a liability into a key advantage for discovering high-quality solutions.
title Optimizing Decoding Paths in Masked Diffusion Models by Quantifying Uncertainty
topic Computation and Language
Artificial Intelligence
Machine Learning
url https://arxiv.org/abs/2512.21336