AI World Models Suffer Instruction Leakage in Spatial Grounding

Compact world models, which use language goals to ground spatial relations like "put the red block left of the blue block," often rely on explicit reference anchors. However, new research reveals a critical flaw: "instruction leakage." This occurs when a goal-conditioned predictor achieves high accuracy (e.g., 90%) not by genuinely perceiving the spatial relation, but by simply transcribing the instruction itself.The study demonstrates that withholding the goal causes accuracy to plummet, and counterfactual instructions lead the predicted anchors to follow the false instruction 94.5% of the time. This confound is characterized as instruction leakage when the scored quantity is directly derivable from the instruction and largely independent of other inputs. This issue was observed in tabletop tasks and the BabyAI benchmark, but not in a Language-Table forward-dynamics model unless the instruction was augmented to name the direction.The diagnosis suggests a clear remedy: the goal should be kept separate from the model's dynamics, belonging instead to the planner's cost function. By supervising the "read path" instead, genuine, instruction-independent grounding can be recovered, achieving high accuracy (88%) both with and without the goal. This detection protocol and fix are applicable to any goal-conditioned world model where the instruction names the quantity being scored.