The Year in AI Papers
The year/Independent research

Paper 2510.11052

Latent Refinement Decoding: Enhancing Diffusion-Based Language Models by Refining Belief States

Qinglin Zhu, Yizhen Yao, Runcong Zhao, Yanzheng Xiang, Amrutha Saseendran, Chen Jin, Philip Teare, Bin Liang, Yulan He, Lin Gui

Published
Oct 2025
Research lab
Independent
Citations
9
GitHub
Not linked

01 In brief

Summary

The paper introduces Latent Refinement Decoding (LRD), a two-stage decoding framework for diffusion-based language models (dLLMs) that addresses information loss and premature commitment in existing hard-assignment methods.

LRD operates in two phases: Phase 1 (Latent Refinement) performs distribution-preserving updates in embedding space by mixing [MASK] embeddings with entropy-normalized top-p token embeddings, allowing the model to establish globally coherent beliefs.

Phase 2 (Predictive Feedback Loop) progressively finalizes confident tokens while retaining uncertain ones in soft form, with KL-divergence dynamics providing a criterion for convergence and early stopping.

Experiments on coding (HumanEval, MBPP) and reasoning (GSM8K, MATH500) benchmarks with LLaDA and Dream models show consistent accuracy improvements (e.g., +6.3 on HumanEval, +3.8 on MATH500) and speedups up to 10.6×.

Ablations confirm that both components contribute, with mixed embeddings being critical for accuracy and early stopping being the main accelerator.

02 From the paper

Abstract

Autoregressive (AR) models remain the standard for natural language generation but still suffer from high latency due to strictly sequential decoding. Recent diffusion-inspired approaches, such as LlaDA and Dream, mitigate this by generating in parallel, yet they suffer from two core limitations: information loss, as predictive distributions for non-finalized tokens are discarded at each step, and premature commitment, where local decisions are made without sufficient global coordination. We introduce Latent Refinement Decoding (LRD), a two-stage framework with Latent Refinement and a Predictive Feedback Loop. The first stage maintains masked positions as distributional mixtures of predicted tokens and the mask embedding, allowing the model to establish more globally consistent beliefs. The second stage progressively finalizes confident tokens while retaining uncertain ones for iterative feedback. KL-divergence dynamics provide a principled and reliable criterion for convergence and early stopping. Experiments across coding (HumanEval +6.3, MBPP +2.6) and reasoning (GSM8K +2.9, MATH500 +3.8) show that LRD improves accuracy while delivering speedups of up to 10.6x, making it a strong and versatile alternative for parallel sequence generation.