Publications

Preprints and Manuscripts

^† Corresponding author.

Refuse and Recompute: What a Classical Planner Fixes in LLM Agents

Xincheng Tan, Ke Wang, Minfeng Zhu^†

Manuscript, 2026

LLM agents that act over long horizons under partial observability lose mostly in two ways: an action after which the goal is unreachable, and acting without knowing what must still happen. A classical planner addresses both from the current state, with a guard that refuses such actions and by recomputing the remaining subgoals.

NSDL: Neuro-Symbolic Division of Labor for Multi-Robot Task Planning

Xincheng Tan, Ke Wang, Yining Di, Minfeng Zhu^†

arXiv:2611.16671, 2026 · Under review at ICRA 2027

A small vision-language model judges the scene and writes a plan sketch; a deterministic solver expands it into an executable joint plan for a heterogeneous robot team. The 7B model reaches 97.0% out-of-distribution accuracy on VIKI-L2.

arXiv Code Models LOCUS-Bench