A goal-conditioned world model scored 0.90 accuracy on spatial-relation readouts — then crumbled to 0.27, pure chance, the moment researchers hid the goal [S1]. That 63-point collapse, reported in a new arXiv preprint from Stony Brook University and Westlake University, exposes a flaw that runs deeper than one model: the system was never perceiving space at all. It was copying the answer from the question. The paper doesn't just diagnose the problem — it offers a fix that recovers 0.88 genuine accuracy. But the implications for anyone building goal-conditioned AI agents are uncomfortable.

The 0.90 that wasn't

The paper, listed as arXiv:2607.06925v1 and authored by Yufeng Wang, Lu Wei, and Haibin Ling [P2], sets up a deceptively simple test. A world model — the internal simulator an AI agent uses to predict what happens next — is trained to reason about spatial relations: is object A to the left of object B? The goal-conditioned version, which receives the task instruction as input, hits 0.90 relation-readout accuracy. By any standard benchmark, that's a strong result [S1].

Except it's hollow. The authors call it "instruction transcription, not perception" [S1]. When they withheld the goal — simply removed the instruction from the input — accuracy collapsed from 0.90 to 0.27 across three random seeds [S1]. The model wasn't looking at the scene. It was reading the answer off the prompt.

The counterfactual test is even more damning. When researchers fed the model a deliberately false instruction — one that contradicted the actual scene — the predicted spatial anchors followed the false instruction 94.5% of the time and matched the true scene just 2.3% of the time, across 256 samples [S1]. The model believed the lie because the lie was all it ever listened to.

How the cheat works

The authors formalise the problem as "instruction leakage" — not to be confused with conventional training-data leakage. Here, leakage happens when the scored quantity (the spatial relation the model is asked to predict) is directly transcribable from the instruction itself, and is essentially independent of how predictive the non-instruction inputs are [S1]. In plain terms: if the instruction says "put the red block to the left of the blue block," and the model is scored on whether red is left of blue, it can ace the test by parroting the instruction without ever processing the visual scene.

This isn't a hypothetical. The authors found that both their tabletop benchmark and the external BabyAI benchmark — a widely used grid-world environment for training language-conditioned agents — exhibit this leakage [S1]. The flaw is baked into the evaluation, not just one model.

One finding narrows the scope. A Language-Table forward-dynamics world model whose instructions named only the referents (the objects) did not leak — until the instruction was augmented to also name the direction (left, right, above) [S1]. Leakage appears precisely when the instruction spells out the answer. The authors also tested whether weakening the action input — degrading the quality of the robot's movement signal — would increase leakage, as a "predictor competition" theory might predict. It didn't. Degrading the action never increased leakage [S1], suggesting the problem is structural, not a side effect of input quality.

What it means

The core insight is uncomfortable for anyone evaluating AI agents. A model can post benchmark numbers that look genuine — 0.90 accuracy, a clean upward curve, consistent across seeds — while doing zero spatial reasoning. The instruction is doing the work; the model is a sophisticated echo chamber. This matters because world models are the engine behind planning agents: systems that simulate future states to decide what action to take next. If the world model can't actually perceive the scene, the agent's plans are built on sand.

The fix the authors propose is elegantly simple: strip the goal out of the dynamics prediction entirely, and instead supervise the read path — the mechanism that extracts spatial relations from the model's internal state [S1]. With the goal removed from the dynamics, the model can no longer cheat by transcribing the instruction. It has to look at the scene. The result: 0.88 accuracy, identical whether the goal is present or withheld [S1]. That identity — same score with and without the goal — is the proof that the grounding is now genuine.

The authors note that their detection protocol and remedy apply to any goal-conditioned world model whose instruction names the scored quantity [S1]. That's a broad class. If instruction leakage is widespread — and the paper suggests it is — some models may be scoring higher than they deserve.

What it means for business

For a two-person robotics startup evaluating off-the-shelf world models, the lesson is concrete: don't trust benchmark accuracy alone. If the model's instruction names the quantity being scored — and most language-conditioned benchmarks do — the headline number may reflect prompt-copying, not perception. The paper's detection protocol is straightforward to run: withhold the goal at inference time and check whether accuracy holds. If it drops to chance, the model is leaking.

For a suburban automation shop building pick-and-place agents, the fix — goal-free dynamics with a supervised read path — is an architectural choice, not a product you can buy today. But it signals a design principle: keep the task instruction out of the prediction loop. Let the world model predict the world; let a separate module extract relations. That separation makes the system auditable — you can test whether the model actually sees the scene, because the scene is all it has.

For teams benchmarking agents on BabyAI or similar environments, the paper is a warning that the evaluation surface itself may be compromised. Re-examining whether instructions leak answers into the scored quantity is now a necessary step before trusting any accuracy figure.

What we don't know yet

This is a single arXiv preprint, v1, with no peer review or independent replication cited [S1]. All statistics — the 0.90 collapse, the 94.5% counterfactual follow rate, the 0.88 recovery — are self-reported by the authors. The 0.88 recovery figure is specific to their proposed architecture and may not generalise to other world-model implementations, model sizes, or task domains.

The paper tests leakage on a tabletop benchmark and BabyAI, both relatively constrained environments. Whether instruction leakage appears at scale in larger, more complex world models — the kind being deployed in production agents — remains an open question. The Language-Table result suggests leakage depends on instruction design, but the full taxonomy of which instruction formats leak and which don't hasn't been mapped.

The next concrete event to watch: peer review and community replication. If independent labs reproduce the 0.90-to-0.27 collapse on BabyAI, instruction leakage moves from a single-paper finding to a known structural risk. Until then, treat any goal-conditioned world model benchmark with the goal visible as a number that might be telling you less than it seems.

Sources: [S1] arXiv:2607.06925v1, "Grounding Spatial Relations in a Compact World Model: Instruction Leakage and a Goal-Free Dynamics Fix," Wang, Wei, Ling (2026). [P2] Full HTML, arxiv.org/html/2607.06925.

If this story changed how you read AI benchmark numbers, subscribe — the next one might change how you read the next one.

Sources

Generated from an audited evidence pack with primary-source research. Social-media items are discussion signals, not verified facts. Nothing here is financial, legal or medical advice.