When Does Flow Matching Help Deterministic Multimodal Prediction: A Spatial Transcriptomics Study

Flow matching is increasingly used beyond open-ended generation, including prediction tasks whose outputs are effectively deterministic, yet it remains unclear why flow matching helps when diversity is not the objective. We study this question in histology-conditioned spatial transcriptomics (ST) prediction, a challenging sparse high-dimensional multimodal task where direct regression with pathology foundation features is already competitive. We disentangle several possible explanations and find that the gain does not come from simply adding a flow objective, injecting noisy target-side states, or increasing architectural complexity. We instead identify intermediate flow states as paired, time-indexed target-side signals: they contain condition-aligned information about the target, but are mixed with source noise and residual variation. Their benefit emerges only when this information is selectively routed into histology representations. Guided by this mechanism, we instantiate conditional flow matching with a simple gated self-attention to learn interactions between histology tokens and noisy ST states. Across high-resolution ST datasets, the model improves spatial correlation over regression and flow-based baselines, with stable convergence and favorable scaling under larger and sparser gene panels. These results establish a design principle for deterministic multimodal prediction: flow matching helps when intermediate target states are treated as structured cross-modal supervision and selectively integrated into conditional representations. Code is available at https://anonymous.4open.science/r/Understanding-When-Flow-Matching-Helps-Deterministic-Multimodal-Prediction-B426/.