Reproduce “MemWM: Memory-Augmented Text-Based World Model” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yujun Wang, Tao Zhang, Jinhe Bi, Aniri, Wenxuan Ye, Boliang Liu, Sikuan Yan, Shuning Wang and 4 more
This paper introduces MEMWM, a memory-augmented text-based world model, along with a Structured State Fidelity (SSF) metric, to enhance AI agent performance in interactive text environments. MEMWM leverages curated world memory and policy-side guidance to improve the accuracy of imagined future states, leading to increased task success across multiple benchmarks without retraining the agent's policy.
Key insights
- Augmenting text-based world models with externalized "world memory" effectively mitigates systematic factual and rule-based prediction errors, improving the trustworthiness of imagined future states.
- A specialized metric like Structured State Fidelity (SSF) is crucial for accurately assessing the preservation of behavior-critical information in generated states, overcoming limitations of general text similarity scores.
- Significant improvements in agent task performance can be achieved by enhancing world models and providing retrieval-time policy guidance, even when the underlying policy model remains frozen.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.