MemPlan: Memory-Conditioned PDDL Planning for Partially Observable Text Environments
This work proposes MemPlan, which recasts large language models from long-horizon planners into constrained interfaces between language and symbolic planning, reconstructing the PDDL problem at each step from positive and negative memory. Its novelty lies in memory-guided ranking of candidate hypotheses and exclusion constraints from refuted assumptions; it is relevant for reducing online token cost and improving task completion in text environments.
01 ABSTRACT
Addressing the high cost and hallucination-prone nature of LLM long-horizon planners, the paper proposes keeping the PDDL domain fixed while reconstructing the PDDL problem each step from parsed observations, candidate hypotheses, and interaction memory. The authors report that on TextWorld, ALFWorld, and Robotouille, MemPlan improves task completion and step efficiency over language-only and PDDL-aware baselines while substantially reducing online token cost; these are the authors' reported results and require independent replication.
02 KEY FINDINGS
- MemPlan fixes the PDDL domain and rebuilds the PDDL problem each step from parsed observations, candidate hypotheses, and interaction memory, forming a memory-conditioned planning loop.
- Positive graph memory assigns relevance-based action costs to candidate hypotheses for missing objects, locations, or conditions, while negative memory compiles refuted assumptions into exclusion constraints.
- A unified schema-conditioned language–symbol interface, trained offline with supervised fine-tuning, handles observation-to-fact parsing and symbolic-action-to-command grounding, and is reused across TextWorld task types.
AI GENERATED SUMMARY / DISCOVERED BY ARXIV CS.CL