Abstract
<title>Abstract</title> <p>Vision–language models can recognize familiar objects while failing to bind familiar attributes in new relational combinations. We ask whether this failure depends not only on what a model sees, but on when it sees relational language. A small cross-modal transformer was trained from scratch on procedurally rendered colored shapes. Paired Early and Late runs had the same initialization, the same indexed image–caption examples, the same hard negatives, and the same number of updates; only the position of the relationally informative and spatially uninformative blocks changed. An identical 200-update interleaved tail tested persistence. The primary test used reciprocal 2 × 2 groups: each of two same-word captions was correct for one image and the foil for its crossed-binding counterpart. This construction cancels any caption-only scoring term while testing conjunctions absent from relational training but present in the control block. Under constant learning rate, Δ = −0.121 (95% bootstrap CI [−0.363, 0.123], exact 𝑝 = 0.398, 𝑑𝑧 = −0.32) for the terminal held-out binding margin. Secondary learning-rate, reversed-stream, and onset analyses did not recover an early advantage. Cross-attention interventions then tested endpoint mechanism. A whole cross-attention layer was selected with separate discovery seeds and semantic validation support, zero-ablated on an untouched test split, and transplanted from seed-matched early models into late models with group-deranged-donor and wrong-layer controls. The intervention pattern did not provide convergent evidence that the selected cross-attention layer specifically carried the behavioral difference. These results do not support an early-exposure advantage in this setting. The scope is a synthetic model and a post-training causal test, not a claim about biological maturation, developmental mediation, or large pretrained systems.</p>