CForce: Boosting Parallel Decoding for dLLMs via Consistency Forcing

2026-08-13

AI Summary

This study addresses the error propagation issue caused by unreliable early predictions during high-parallelism decoding in diffusion large language models. We propose Consistency-Enforced Distillation, which aligns early and late masked predictions through a confidence-adaptive KL divergence objective within a unified decoding framework. Augmented with self-rollback trajectory training, this approach is theoretically proven to minimize early prediction errors effectively. Experiments on the LLaDA model demonstrate that our method significantly suppresses error propagation and substantially improves the speed-quality trade-off under high parallelism budgets. Consequently, this work enhances both inference efficiency and generation stability for diffusion-based language modeling, offering a robust solution to the challenges of accelerated decoding without compromising output fidelity.

Abstract

Diffusion large language models (dLLMs) accelerate language generation by predicting multiple masks in a single forward pass. However, existing dLLMs can suffer from unreliable predictions in early denoising stages under aggressive parallelism strategies, leading to errors that can propagate to later stages. To tackle this issue, we present Consistency Forcing (CForce) for dLLMs, a distillation method to force the mask predictions of early stages to align with those of later stages. CForce trains the model on pre-collected self-rollout trajectories, thereby improving training-inference alignment. We introduce Confidence Adaptive KL Divergence as a distillation objective to conjoin the merits of forward and reverse KL. We further provide a theoretical analysis for the consistency objective to explain why CForce can approximately minimize the prediction error of early stages. Critically, the same formulation applies to both mask-to-token decoding and edit-capable decoding; in the edit-capable case, later token-to-token refinements provide additional supervision for earlier masked-state predictions. Experiments on non-edit and edit-capable LLaDA models show improved speed-quality trade-offs, especially under high-parallelism decoding budgets. Code is available at: https://github.com/inclusionAI/dFactory.

Problem

• Diffusion Large Language Models
• Parallel Decoding
• Error Propagation
• Early Denoising Stages

Innovation

• Consistency Forcing
• Diffusion Large Language Models
• Confidence Adaptive KL Divergence
• Parallel Decoding
• Self-rollout Trajectories