Reproduce “Learning Dynamic Belief Graphs to Generalize on Text-Based Games” 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
Ashutosh Adhikari, Xingdi Yuan, Marc-Alexandre Côté, Mikuláš Zelinka, Marc-Antoine Rondeau, Romain Laroche, Pascal Poupart, Jian Tang and 2 more
The Graph-Aided Transformer Agent (GATA) learns dynamic, continuous belief graphs directly from raw text observations to improve decision-making in text-based games. This approach enhances agent performance by an average of 24.2% over text-based baselines and demonstrates improved generalization across diverse game configurations.
Key insights
- Dynamically learned, continuous graph-structured belief states provide a valuable inductive bias for language-based agents, leading to better policies and enhanced generalization in partially observable environments.
- Self-supervised pre-training without explicit ground-truth graph supervision can effectively enable models to infer structured relational information directly from raw textual game observations.
- Combining learned graph representations with direct text observations via attention mechanisms allows for grounded action scoring, mitigating errors in the learned graph and focusing on currently observable entities.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.