MemWM: Memory-Augmented Text-Based World Model Improves Planning Accuracy
A recent research article presents MemWM, a text-based world model enhanced with memory, aimed at boosting AI agents' planning capabilities by refining next-state predictions. This model tackles systematic inaccuracies in world models, such as neglecting essential task details, distorting product characteristics, and misapplying transition rules. MemWM features a specialized memory repository, known as 'world memory,' which holds transition rules, state caches, and challenging-to-predict information, guiding next-state generation based on this memory. Additionally, the study introduces Structured State Fidelity (SSF), a novel evaluation metric that assesses predicted states against specific benchmark facts. Findings indicate that memory-augmented training enhances SSF by as much as 206.3% over standard fine-tuning (SFT). The research can be found on arXiv with the identifier 2608.07107.
Key facts
- MemWM is a memory-augmented text-based world model.
- It uses a curated memory bank of transition rules, state caches, and hard-to-predict facts.
- Structured State Fidelity (SSF) is introduced to evaluate factual state preservation.
- Memory-augmented training improves SSF by up to 206.3% compared to SFT.
- In planning, the policy model is frozen and provided with policy-side world skills.
- The paper is available on arXiv with ID 2608.07107.
- The model addresses systematic prediction errors in world models.
- The research focuses on improving next-state predictions for agent planning.
Entities
Institutions
- arXiv