Grounding Inferred Relationships in Complex World Models with Continual Reasoning

Preliminary Abstract

Language models struggle to infer and reason about implicit relationships in complex world models, significantly limiting their decision-making capabilities. Benchmarks like ARC-AGI and DABStep illustrate these challenges, where state-of-the-art models achieve only 55% and 16.4% accuracy, respectively. We propose a Continual Reasoning framework to ground inferred relationships at inference time, enabling models to learn premises dynamically without a predefined domain-specific language. Our approach utilizes an external memory structured as a dynamic hypothesis graph, allowing iterative refinement of relationships through interaction. We hypothesize that defining entity containers within a structured memory can optimize the hypothesis search space, making inference in complex world models more efficient. This work aims to enhance AI-driven scientific discovery and decision-making in real-world, data-driven environments.