GAVEL: Graph World Models for Verified and Efficient Long-Horizon LLM Task Planning
GAVEL uses a graph world model to check, fix, and reorder LLM robot plans before execution.
The framework models object relations, action preconditions and effects, and probabilistic beliefs about unobserved object locations. It repairs errors when the graph model supplies a direct correction, leaving LLM replanning for cases that need semantic reasoning. On BEHAVIOR-1K, the paper reports Qwen3-8B success rising from 41.2% to 91.8% on single long-horizon tasks and from 19.9% to 92.6% on multi-task instructions. Distributional belief reasoning cut travel distance by about 5.4% versus a static variant. HF Daily Papers' note
The framework models object relations, action preconditions and effects, and probabilistic beliefs about unobserved object locations. It repairs errors when the graph model supplies a direct correction, leaving LLM replanning for cases that need semantic reasoning. On BEHAVIOR-1K, the paper reports Qwen3-8B success rising from 41.2% to 91.8% on single long-horizon tasks and from 19.9% to 92.6% on multi-task instructions. Distributional belief reasoning cut travel distance by about 5.4% versus a static variant. HF Daily Papers' note
score 5