Localize, Then Reason: Visual Latent Structural Reasoning for Molecular Properties and Edits

Visual Latent Structural Reasoning (VLSR) integrates molecular region localization with property reasoning via an end-to-end framework. The model identifies chemically significant motifs within images and processes them in a compact latent workspace. This architecture achieves 9.6X higher throughput compared to baseline textual-reasoning methods during inference.