What Has a Foundation Model Found? Using Inductive Bias to Probe for World Models

2025-07-09

AI Summary

This work investigates whether foundation models genuinely acquire deep world models—such as fundamental physical laws—or merely rely on task-specific heuristics. Method: We introduce the first “inductive bias probing” framework: synthetic sequential data are generated under hypothesized world models (e.g., Newtonian mechanics), and models are rigorously evaluated via cross-domain generalization tests and quantitative bias analysis to characterize their intrinsic inductive tendencies. Contribution/Results: Experiments reveal that state-of-the-art foundation models, while achieving strong performance within their training distribution, fail to develop inductive biases aligned with real-world physical principles—such as orbital mechanics—exposing a fundamental limitation in their generalization mechanisms. Our framework provides a scalable, interpretable, and theory-grounded paradigm for assessing the cognitive capabilities of large models, bridging formal physics-based assumptions with empirical model evaluation.

Technology Category

• Machine Learning: Deep Neural Architectures and Foundation Models
• Cognitive Modeling & Cognitive Systems: Other Foundations of Cognitive Modeling & Systems
• Humans and AI: Other Foundations of Human Computation & AI

Application Category

• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Web Mining and Content Analysis: Large pretrained models with web data

Abstract

Foundation models are premised on the idea that sequence prediction can uncover deeper domain understanding, much like how Kepler's predictions of planetary motion later led to the discovery of Newtonian mechanics. However, evaluating whether these models truly capture deeper structure remains a challenge. We develop a technique for evaluating foundation models that examines how they adapt to synthetic datasets generated from some postulated world model. Our technique measures whether the foundation model's inductive bias aligns with the world model, and so we refer to it as an inductive bias probe. Across multiple domains, we find that foundation models can excel at their training tasks yet fail to develop inductive biases towards the underlying world model when adapted to new tasks. We particularly find that foundation models trained on orbital trajectories consistently fail to apply Newtonian mechanics when adapted to new physics tasks. Further analysis reveals that these models behave as if they develop task-specific heuristics that fail to generalize.

Problem

• Evaluating if foundation models capture deeper domain structure
• Assessing model adaptation to synthetic world model datasets
• Testing inductive bias alignment with underlying world models

Innovation

• Probing models with synthetic datasets
• Measuring inductive bias alignment
• Evaluating generalization via task adaptation