When Does Explicit View Routing Work? A Controlled Study of Multi-View Graph-Text Alignment

Modern graph-text retrieval models often use multiple heads to align different semantic aspects (like molecular properties or class labels) with graph structures. However, it is unclear whether these models perform true semantic routing—where the model understands the content of the query—or if they simply rely on architectural channelization, where the model responds to the selected head regardless of the input text. This study investigates the conditions under which explicit view routing can be verified.

To distinguish between semantic routing and architectural channelization, the authors implement a controlled version of Multi-View Graph-Text Alignment (MV-GTA). They use deterministic, verifiable text segments and isolated encoders to ensure that information cannot leak across views. The study employs a causal control method: they compare "correct" routing against "deranged" routing, where text segments are mismatched with their intended channels. By using external labels and RDKit descriptors as ground-truth relevance, the researchers can verify if retrieval performance depends on the actual content of the text rather than just the choice of the graph head.

The experiments demonstrate that for label and property views, the models exhibit genuine content-dependent routing; correct routing significantly outperforms deranged training. However, this success is not universal. Topology, in particular, does not consistently specialize across datasets, and the expected graph head does not always outperform the best "wrong" head. While joint models (trained on all views) show higher aggregate retrieval scores than individual specialists, this is descriptive of functional consolidation rather than statistical equivalence. Furthermore, while paraphrase augmentation improves robustness to linguistic variation, it often introduces trade-offs, where gains in out-of-distribution performance come at the cost of canonical retrieval accuracy.