Researchers Expose Critical Flaw in Language-Conditioned World Models: Instruction Leakage Problem

A new arXiv study identifies a fundamental vulnerability in compact world models used for robotic spatial reasoning: models appear to achieve 90% accuracy at understanding instructions like "put the red block left of the blue block," but are actually just transcribing the instruction rather than perceiving the scene. The researchers demonstrate the flaw collapses to 27% accuracy when the goal is withheld and propose a fix—keeping goals out of dynamics prediction while supervising the read path—that recovers genuine spatial grounding.

Why it matters: This research exposes a widespread methodological pitfall in AI systems designed for robot instruction following, affecting how practitioners evaluate and design language-grounded world models across robotics and embodied AI.