Search papers, labs, and topics across Lattice.
This paper investigates the phenomenon of instruction leakage in compact world models that utilize language goals to ground spatial relations. The authors reveal that a goal-conditioned predictor achieves high accuracy through instruction transcription rather than genuine perception, leading to significant drops in performance when the goal is withheld. By proposing a method to decouple the goal from the dynamics, they demonstrate that genuine grounding can be achieved without reliance on the instruction, maintaining accuracy while preventing leakage.
Instruction leakage can lead to misleadingly high accuracy in spatial relation tasks, revealing a critical flaw in goal-conditioned models that could misguide future research.
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 \emph{reference anchors}. We ask 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 \emph{instruction transcription}, not perception. Withholding the goal collapses it to chance ($0.90\!\to\!0.27$, three seeds) and a counterfactual instruction makes the predicted anchors follow the \emph{false} instruction $94.5\%$ of the time (true scene $2.3\%$; $N{=}256$). Tested across three settings and a within-task ablation, our central claim characterizes the confound: \textbf{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 \emph{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 \emph{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.
