Fine-Tuning Masked Diffusion for Provable Self-Correction

A natural desideratum for generative models is self-correction--detecting and revising low-quality tokens at inference. While Masked Diffusion Models (MDMs) have emerged as a promising approach for generative modeling in discrete spaces, their capacity for self-correction remains poorly understood. Prior attempts to incorporate self-correction into MDMs either require overhauling MDM architectures/training or rely on imprecise proxies for token quality, limiting their applicability. Motivated by this, we introduce PRISM--Plug-in Remasking for Inference-time Self-correction of Masked Diffusions--a lightweight, model-agnostic approach that applies to any pretrained MDM. Theoretically, PRISM defines a self-correction loss that provably learns per-token quality scores, without RL or a verifier. These quality scores are computed in the same forward pass with MDM and used to detect low-quality tokens. Empirically, PRISM advances MDM inference across domains and scales: Sudoku; unconditional text (170M); and code with LLaDA (8B). We open-source our codebase in https://github.com/SeunggeunKimkr/PRISM.

Generative models often make mistakes during generation, and a useful ability is self-correction: detecting bad tokens and revising them. We introduce PRISM, a lightweight method that adds this ability to pretrained Masked Diffusion Models without reinforcement learning, external verifiers, or major architectural changes. PRISM learns per-token quality scores, computes them during the model’s normal forward pass, and uses them to remask and regenerate low-quality tokens. Across Sudoku, text generation, and code generation with an 8B model, PRISM improves inference and helps MDMs produce better outputs.