Sliding Window Attention with Auditable Chemical Sinks for Long-Sequence Molecular Transformation Modeling

Long molecular transformation sequences require models to track local bond changes and persistent reaction constraints. Conventional Transformers face high context costs, diluted attention, and chemically invalid predictions. We propose ChemSWA, an efficient and auditable framework for multi-step synthesis modeling. Adaptive Chemical Sliding Window Attention combines local windows with persistent attention sinks to retain reaction states at 𝑂 (𝑊 · 𝐿) computational complexity. A constraint-audited attribution mechanism checks chemical conservation and produces step-wise certificates, while counterfactual calibration reduces error accumulation during rollout. On USPTO-50K, ChemSWA achieves 52.14% Top-1 accuracy and 99.95% chemical validity; it reaches 92.85% Top-1 accuracy on USPTO-MIT. On PaRoutes, it achieves 76.10% Solve@10 with 3.85 s average expansion latency, and on USPTO-Condition it reaches 0.3015 multi-slot accuracy.