DiffuPlan: Editable Planning State for Diffusion Language Model Agents

Key points

• We introduce DiffuPlan, a diffusion search agent that maintains a single editable planning state and reconstructs only the affected steps when later evidence resolves an entity or changes the search route.1
• To support variable-length revisions, we introduce Expansion-Partition Marginalization (EPM), which marginalizes over equivalent expansion paths.1
• We further adapt diffusion trajectory RL to multi-turn search.1
• These results show that dLLMs need not imitate append-only AR agents: revising committed state is a distinct agentic advantage.1

Sources (1)

• [1]DiffuPlan: Editable Planning State for Diffusion Language Model Agents
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 17, 02:30 PM

  Diffusion language models (dLLMs) are emerging as an alternative to autoregressive (AR) language models, yet dLLM agents largely inherit an interaction protocol designed for AR generation: once intermediate state is written, later observations can only be incorporated by appending more text.

  We introduce DiffuPlan, a diffusion search agent that maintains a single editable planning state and reconstructs only the affected steps when later evidence resolves an entity or changes the search route.

Extractive summary: sentences quoted from the sources.