LLMs vs. World Models: A New Framework for AGI

A recent study published on arXiv (2605.23972) claims that large language models (LLMs) face inherent limitations in causal reasoning, maintaining persistent states, and planning over extended periods. This is attributed to a mismatch at the objective level between sequence prediction and reasoning about latent environmental dynamics. The researchers propose a new framework called Latent Dynamics Inference (LDI), which views language and multimodal inputs as partial indicators of underlying transition dynamics. To validate their concept, they introduce Flux, a sequential reasoning environment governed by natural-language rules. As a proof-of-concept, these rules are transformed into a state-transition simulator, demonstrating that structured latent dynamics can occasionally be derived from text. The findings indicate that world models might excel over LLMs in tasks necessitating a profound comprehension of causality.