Generalized Discrete Diffusion with Self-Correction

2026-02-13

AI Summary

Existing self-correction methods for discrete diffusion models are typically introduced during inference or post-training, exhibiting limited generalization and often degrading performance. This work proposes the Self-Correcting Discrete Diffusion (SCDD) model, which, for the first time, integrates an explicit state-transition-based self-correction mechanism directly into pretraining within a discrete-time framework. SCDD eliminates redundant re-masking steps and relies solely on a uniform absorption objective for learning. By simplifying the noise schedule and combining BERT-style pretraining with parallel decoding, the method substantially improves decoding efficiency—demonstrated on GPT-2-scale experiments—while preserving high generation quality.

Abstract

Self-correction is an effective technique for maintaining parallel sampling in discrete diffusion models with minimal performance degradation. Prior work has explored self-correction at inference time or during post-training; however, such approaches often suffer from limited generalization and may impair reasoning performance. GIDD pioneers pretraining-based self-correction via a multi-step BERT-style uniform-absorbing objective. However, GIDD relies on a continuous interpolation-based pipeline with opaque interactions between uniform transitions and absorbing masks, which complicates hyperparameter tuning and hinders practical performance. In this work, we propose a Self-Correcting Discrete Diffusion (SCDD) model to reformulate pretrained self-correction with explicit state transitions and learn directly in discrete time. Our framework also simplifies the training noise schedule, eliminates a redundant remasking step, and relies exclusively on uniform transitions to learn self-correction. Experiments at the GPT-2 scale demonstrate that our method enables more efficient parallel decoding while preserving generation quality.

Problem

• discrete diffusion
• self-correction
• pretraining
• uniform transitions
• parallel sampling

Innovation

• self-correction
• discrete diffusion
• uniform transitions
• parallel decoding
• pretraining