Diffusion language models (dLLMs) are gaining attention as an alternative to autoregressive language models. Yet, according to a new arXiv paper, dLLM agents have largely inherited an interaction protocol designed for autoregressive generation: once intermediate state is written, later steps are constrained by that fixed output.
The paper, titled DiffuPlan: Editable Planning State for Diffusion Language Model Agents, proposes a different design. DiffuPlan gives agents an editable planning state, allowing intermediate decisions to be revisited and changed rather than locked in after being generated.
The motivation is that planning benefits from revision, and diffusion models—which generate by iteratively refining—are naturally suited to this kind of editing. The abstract frames this as a way to move dLLM agents beyond the AR-style protocol they currently use.
Because the abstract is brief, it does not include implementation specifics or evaluation results. The contribution at this stage is conceptual: a planning-state mechanism that treats agent outputs as revisable rather than permanent.