Compact World Model: instruction leakage and goal-free fix

Grounding Spatial Relations in a Compact World Model, a paper by Yufeng Wang, Lu Wei and Haibin Ling submitted 8 Jul 2026, shows that high relation-readout scores can hide a pathological shortcut the authors call "instruction transcription, not perception." The work measures when reference anchors truly reflect scene perception and when they simply reproduce the instruction.

What did the paper find?

The paper finds that goal-conditioned predictors can reach a relation-readout accuracy of 0.90 while not actually perceiving the scene; withholding the instruction collapses accuracy to chance, 0.27 (three seeds). The authors ran counterfactual tests: when given a false instruction the predictor followed the false instruction 94.5% of the time and matched the true scene only 2.3% of the time (N=256). They tested across three settings, ran a within-task ablation, and showed that common benchmarks leak.

Across their evaluations the tabletop benchmark and the external BabyAI benchmark exhibited instruction leakage. A Language-Table forward-dynamics world model did not leak when the instruction only named referents; leakage appeared once the instruction was augmented to name the direction. The authors also report that degrading the action signal never increases leakage, which runs counter to what predictor-competition explanations predict.

How did the authors diagnose and fix the leakage?

They diagnose the confound as instruction leakage: any scored quantity that is directly transcribable from the instruction can be solved by transcription, independent of the non-instruction inputs. The detection protocol includes withholding the goal, using counterfactual instructions, and a within-task ablation to test whether predictions depend on the scene or on the instruction. The counterfactual test produced the 94.5% false-instruction follow rate (true scene 2.3%, N=256), which exposed transcriptional behavior.

The prescribed remedy is architectural: keep the goal out of the dynamics model and treat the goal as planner cost, while supervising the read path (the anchor prediction). Applying this goal-free dynamics fix recovers genuine, instruction-independent grounding with a readout accuracy of 0.88, identical with and without the goal. The paper frames the fix as general: it applies to any goal-conditioned world model where the instruction names the scored quantity.

Why it matters

High readout scores alone can mislead researchers and practitioners into believing a model grounded relations from perception when it has merely learned to transcribe instructions. The paper supplies concrete diagnostics and a concrete engineering fix. That matters for any system that conditions world models on language goals, because unchecked leakage makes evaluations and downstream planners trust signals that do not reflect the environment.

What to watch

Look for follow-up evaluations that apply the detection protocol to additional benchmarks beyond tabletop and BabyAI, and for implementations of the goal-free dynamics pattern in other forward-dynamics architectures. A clear confirmation signal will be reproduction of the paper's key numbers: the 0.90 versus 0.27 collapse when withholding goals, the 94.5% counterfactual follow rate (true scene 2.3%, N=256), and recovery to 0.88 grounding accuracy with the fix.

References and provenance: the findings, numbers and tests above come from the paper "Grounding Spatial Relations in a Compact World Model: Instruction Leakage and a Goal-Free Dynamics Fix," submitted to arXiv on 8 Jul 2026 by Yufeng Wang, Lu Wei and Haibin Ling.