Autoregressive Planning for Diffusion Language Agent

Autoregressive language models form the backbone of modern agentic systems. Diffusion language models (dLLM) offer an alternative to autoregressive (AR) decoding because they can refine many tokens in parallel. However, recent agentic evaluations show that this speed comes with failures in precision and multi-step tool use. We propose commitment-conditioned diffusion agents (CCDA), a hybrid AR--diffusion approach that separates commitments from parallel execution. An autoregressive planner first emits a compact skeleton plan specifying key information; then a diffusion language model based executor denoises the action conditioned on this skeleton. We formulate training-free and trainable versions of CCDA and organize evaluation around agentic tasks, where prior work exposes complementary weaknesses and strengths of dLLM agents.