Where Should Diffusion Enter a Language Model? Geometry-Guided Hidden-State Replacement

Injin Kong, Hyoungjoon Lee, Yohan Jo

Keywords: diffusion language models, transformer replacement, hidden-state diffusion, representation geometry, layer selection, pretrained language models

TL;DR: We show that continuous diffusion works better inside a pretrained transformer when it reconstructs a geometry-selected hidden state rather than tokens or embeddings.

Abstract: Continuous diffusion language models lag behind autoregressive transformers, partly because diffusion is applied in spaces poorly suited to language denoising and token recovery. We propose DiHAL, a geometry-guided diffusion-transformer hybrid that asks where diffusion should enter a pretrained transformer. By scoring layers with geometry-based proxies, DiHAL selects a diffusion-friendly hidden-state interface and replaces the lower transformer prefix with a diffusion bridge, retaining the upper layers and original LM head to avoid direct continuous-to-discrete recovery. Experiments on 8B-scale backbones show that the geometry score predicts effective shallow insertion layers under a fixed bridge-training protocol, and that hidden-state recovery improves over continuous diffusion baselines in a diagnostic comparison matching the training budget. These results suggest that hidden-state geometry helps identify where diffusion-based replacement is feasible inside pretrained language models.

Submission Number: 77