Gilad Turok

Gilad Turok, Chris De Sa, Volodymyr Kuleshov

Masked diffusion models (MDMs) generate text by iteratively selecting positions to unmask and then predicting tokens at those positions. Yet MDMs lack proper perplexity evaluation: the ELBO is a loose bound on likelihood under the training distribution, not the test-time distribution, while generative perplexity requires a biased external model and ignores diversity. To address this, we introduce the \textsc{DUEL} framework, which formalizes \emph{deterministic} position selection, unifying leading MDM sampling strategies. We prove \textbf{\textsc{DUEL} admits \emph{exact} likelihood computation} via a simple algorithm, evaluated under the same position selection used at test time. This \textbf{gives MDMs proper perplexity for the first time} -- the natural analogue of autoregressive perplexity. With proper perplexity in hand, we revisit key questions about MDMs. \textbf{MDMs are substantially better than previously thought}: the MDM-autoregressive perplexity gap shrinks by up to 32\% on in-domain data and 82\% on zero-shot benchmarks. \textsc{DUEL} enables the first principled comparison of fast, parallel samplers across compute budgets -- an analysis impossible with the ELBO and unreliable with generative perplexity -- identifying probability margin \citep{kim2025train} as a strong default. Finally, oracle search over position orderings reveals MDMs can far surpass autoregressive models -- achieving 36.47 vs.\ 52.11 perplexity on AG News -- demonstrating the ceiling of MDM performance has not yet been reached.

Yair Schiff, Omer Belhasin, Roy Uziel et al.

Masked diffusion models (MDMs) have emerged as a promising alternative to autoregressive models, enabling parallel token generation while achieving competitive performance. Despite these advantages, MDMs face a fundamental limitation: once tokens are unmasked, they remain fixed, leading to error accumulation and ultimately degrading sample quality. We address this by proposing a framework that trains a model to perform both unmasking and correction. By reusing outputs from the MDM denoising network as inputs for corrector training, we train a model to recover from potential mistakes. During generation we apply additional corrective refinement steps between unmasking ones in order to change decoded tokens and improve outputs. We name our training and sampling method Progressive Self-Correction (ProSeCo) for its unique ability to iteratively refine an entire sequence, including already generated tokens. We conduct extensive experimental validation across multiple conditional and unconditional tasks, demonstrating that ProSeCo yields better quality-efficiency trade-offs (up to ~2-3x faster sampling) and enables inference-time compute scaling to further increase sample quality beyond standard MDMs (up to ~1.3x improvement on benchmarks).

Guanghan Wang, Yair Schiff, Gilad Turok et al.

While diffusion language models (DLMs) have achieved competitive performance in text generation, improving their reasoning ability with reinforcement learning remains an active research area. Here, we introduce d2, a reasoning framework tailored for masked DLMs. Central to our framework is a new policy gradient algorithm that relies on properties of masking to accurately estimate the likelihoods of sampling trajectories. Our estimators trade off computation for approximation accuracy in an analytically tractable manner, and are particularly effective for DLMs that support any-order likelihood estimation. We characterize and study this property in popular DLMs and show that it is key for efficient diffusion-based reasoning. Empirically, d2 significantly improves over previous diffusion reasoning frameworks using only RL (without relying on supervised fine-tuning), and sets a new state-of-the-art performance for DLMs on logical reasoning tasks (Countdown and Sudoku) and math reasoning benchmarks (GSM8K and MATH500).