Rethinking Token Prediction: Tree-Structured Diffusion Language Model

arXiv cs.CL / 4/7/2026

💬 OpinionIdeas & Deep AnalysisModels & Research

共有:

Key Points

The paper argues that discrete diffusion language models are currently inefficient because the full-vocabulary token prediction head consumes a large share of parameters and dominates peak GPU memory.
It proposes a tree-structured diffusion approach that replaces full-vocabulary classification with predictions over a vocabulary tree using ancestor-based latent states, drastically reducing classification dimensionality.
By making the prediction head nearly negligible, the method reallocates capacity to deepen attention blocks while keeping the overall parameter budget fixed.
Experiments report a 50% reduction in peak GPU memory usage while matching state-of-the-art perplexity results for discrete diffusion language models.
Overall, the work reframes token prediction as a structured factorization problem, aiming to make diffusion-based LLM training more practical under tight hardware limits.

Abstract

Discrete diffusion language models have emerged as a competitive alternative to auto-regressive language models, but training them efficiently under limited parameter and memory budgets remains challenging. Modern architectures are predominantly based on a full-vocabulary token prediction layer, which accounts for a substantial fraction of model parameters (e.g., more than 20% in small scale DiT-style designs) and often dominates peak GPU memory usage. This leads to inefficient use of both parameters and memory under constrained training resources. To address this issue, we revisit the necessity of explicit full-vocabulary prediction, and instead exploit the inherent structure among tokens to build a tree-structured diffusion language model. Specifically, we model the diffusion process with intermediate latent states corresponding to a token's ancestor nodes in a pre-constructed vocabulary tree. This tree-structured factorization exponentially reduces the classification dimensionality, makes the prediction head negligible in size, and enables reallocation of parameters to deepen the attention blocks. Empirically, under the same parameter budget, our method reduces peak GPU memory usage by half while matching the perplexity performance of state-of-the-art discrete diffusion language models.