Abstract
Compact world models that condition on a language goal promise to ground relations such as "put the red block left of the blue block" using a sparse set of explicit reference anchors. We investigate when such references actually ground a relation and identify a trap: a goal-conditioned predictor reaches a striking $0.90$ relation-readout accuracy, yet this is instruction transcription, not perception.
Withholding the goal collapses it to chance ($0.27$, three seeds) and a counterfactual instruction makes the predicted anchors follow the false instruction $94.5\,\text{perthousand}$ of the time (true scene $2.3\,\text{perthousand}$; $N=256$).
Tested across three settings and a within-task ablation, our central claim characterizes the confound: instruction leakage occurs when the scored quantity is transcribable from the instruction (when the instruction names the answer) and is essentially independent of how predictive the non-instruction inputs are. Our tabletop and the external BabyAI benchmark leak, whereas a Language-Table forward-dynamics world model whose instruction names referents does not, until the instruction is augmented to name the direction; and degrading the action never increases leakage, the opposite of what predictor-competition predicts.
The diagnosis prescribes the fix: keep the goal out of the dynamics (it belongs to the planner's cost) and supervise the read path, recovering genuine, instruction-independent grounding ($0.88$, identical with and without the goal). The detection protocol and remedy apply to any goal-conditioned world model whose instruction names the scored quantity.
Blogger's Review: This paper delves into the issue of instruction leakage in compact world models, emphasizing the distinction between instruction and perception while proposing effective solutions. This research is significant for building more accurate language understanding models, making it worthy of attention.