Planner Aware Path Learning in Diffusion Language Models Training

ICLROral2026

Authors
Fred Zhangzhi Peng, Zachary Bezemek, Jarrid Rector-Brooks, Shuibai Zhang, Michael M. Bronstein, Anru Zhang, Alexander Tong, Joey Bose
Affiliation
Duke University
Venue
ICLR 2026
Track
Oral

TL;DR

We propose Planner Aware Path Learning (PAPL), a simple planner-aligned training method for Diffusion Language Models that resolves the training–inference mismatch and consistently improves generation quality.

Opening excerpt from the authors’ abstract. source

Read the paper

Topics

language model diffusion

← All ICLR 2026 Oral papers · Browse the whole archive