Optimizing Decoding Paths in Masked Diffusion Models by Quantifying Uncertainty

2025-12-24

AI Summary

Masked diffusion models (MDMs) suffer from unstable non-autoregressive generation quality due to suboptimal decoding path selection. This work attributes the performance gap to accumulated predictive uncertainty during denoising and formally defines *denoising entropy*—a computable, intrinsic metric for assessing uncertainty along the denoising trajectory. Building on this insight, we propose two key innovations: (1) a posterior-probability-based optimal path selection algorithm that explicitly maximizes path likelihood; and (2) a real-time entropy-guided dynamic denoising scheduling mechanism that actively leverages uncertainty as an optimization signal to adaptively refine denoising steps. Evaluated on challenging benchmarks spanning reasoning, planning, and code generation, our approach significantly improves both accuracy and output stability over prior MDM methods. Results demonstrate that entropy-guided denoising is not only effective for enhancing non-autoregressive generation quality but also broadly applicable across diverse structured prediction tasks.

Technology Category

• Machine Learning: Calibration & Uncertainty Quantification
• Computer Vision: Diffusion Models for Vision
• Reasoning under Uncertainty: Sequential Decision Making

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Abstract

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.

Problem

• Quantifies predictive uncertainty in masked diffusion models' decoding paths.
• Optimizes decoding order to enhance generation quality and accuracy.
• Transforms uncertainty into an advantage for discovering high-quality solutions.

Innovation

• Introducing Denoising Entropy to quantify predictive uncertainty
• Proposing post-hoc selection and real-time guidance algorithms
• Using entropy-guided methods to improve generation quality