Why We Need World Models for AGI: Where LLMs Fail and How World Models May Outperform

In three linesarXiv paper arguing LLMs fail at causal reasoning and long-horizon planning due to lack of world models. Authors introduce Latent Dynamics Inference (LDI) and Flux, a sequential reasoning environment specified in natural language. RL agents with explicit latent state access achieve 79% win rate vs 11% for LLMs, revealing failures in persistent state tracking.

Summary generated by Claude — human-verified