Evaluating the World Model Implicit in a Generative Model
This addresses the challenge of assessing world model learning in AI for researchers, revealing limitations in current models and offering new evaluation methods, though it is incremental in refining diagnostic tools.
The paper tackles the problem of evaluating whether generative models implicitly learn coherent world models by proposing new metrics based on the Myhill-Nerode theorem, applied to domains like game-playing and navigation, and finds that models perform well on existing diagnostics but show significant incoherence, leading to fragility in related tasks.
Recent work suggests that large language models may implicitly learn world models. How should we assess this possibility? We formalize this question for the case where the underlying reality is governed by a deterministic finite automaton. This includes problems as diverse as simple logical reasoning, geographic navigation, game-playing, and chemistry. We propose new evaluation metrics for world model recovery inspired by the classic Myhill-Nerode theorem from language theory. We illustrate their utility in three domains: game playing, logic puzzles, and navigation. In all domains, the generative models we consider do well on existing diagnostics for assessing world models, but our evaluation metrics reveal their world models to be far less coherent than they appear. Such incoherence creates fragility: using a generative model to solve related but subtly different tasks can lead to failures. Building generative models that meaningfully capture the underlying logic of the domains they model would be immensely valuable; our results suggest new ways to assess how close a given model is to that goal.