Diffusion language models reveal uncertainty from generation dynamics
From Denoising Dynamics to Endpoint Uncertainty in Diffusion Language Models
Artificial Intelligence
Summary
Generating text with AI models can be unpredictable, and it usually requires producing many complete answers to understand this uncertainty. The authors studied a special kind of language model that creates text step-by-step by gradually refining it, called diffusion models. They found that the way the text changes within a single generation run can tell us how uncertain the final answers will be across many runs. This insight means you could estimate uncertainty with fewer full text generations, saving time and computing.
What this means in practice
- •For natural language processing engineers: Estimate text generation uncertainty with fewer full runs by analyzing partial output changes during diffusion.
- •For ai system reliability teams: Use denoising step dynamics to better predict when repeated model outputs will agree or differ, improving confidence scoring.
Authors
Julio C. Amador Diaz Lopez
Abstract
Predictive uncertainty in stochastic language-model inference can be characterized by repeatedly sampling an input and measuring the distribution of outputs. This requires multiple complete generations. Masked diffusion language models expose another source of information: the evolution of the output during denoising. We ask whether convergence within diffusion trajectories is associated with uncertainty across their final outputs, or endpoints. We study four masked diffusion language models on four reasoning benchmarks, collecting 20 generations per question and 32,000 denoising trajectories in total. Within each trajectory, we measure within-run concentration as the area under the curve describing stable resolution of the eventual output. Across repeated trajectories, we measure endpoint-distribution uncertainty. Stronger within-run concentration is associated with lower endpoint entropy after accounting for model and dataset ($β=-0.222$, $p<10^{-25}$). This association persists for multi-token outputs and when reweighting positive-entropy cases. The relationship is asymmetric: high within-run concentration almost exclusively accompanies concentrated endpoint distributions, whereas lower concentration occurs under both concentrated and competing endpoint regimes. This relationship strengthens as denoising proceeds. Features from a single denoising run predict whether independent held-out runs will concentrate on a common endpoint or produce competing endpoints. These results connect dynamics within individual denoising runs to variability across repeated inference. Within-trajectory dynamics contain information about whether independent repetitions will concentrate on a common endpoint or produce competing endpoints, providing a route towards estimating endpoint-distribution properties with fewer complete generations.