Predicting the potential-dependent optical response of electrochromic polymers (ECPs) from molecular structure remains a major challenge because conventional quantum chemical methods are computationally prohibitive for charged conjugated systems with extensive conformational diversity and complex solvent and solid-state interactions. Here, we develop a multi-state neural network framework for predicting continuous UV–Vis–NIR absorption spectra across multiple oxidation states, enabling data-driven modeling of polymer spectroelectrochemical behavior. To overcome data scarcity, we establish a hierarchical database that transfers knowledge from data-rich small-molecule spectroscopic records to neutral and charged conjugated polymers, progressively incorporating in-house experimental measurements. Quantum-mechanically derived electronic and structural descriptors encode three-dimensional geometry, molecular topology, charge distribution, and long-range interactions to improve predictive accuracy. The framework is further integrated with an automated robotic synthesis and characterization platform, enabling closed-loop active learning in which model-guided experiments are selected to maximize information gain and iteratively refine predictions. By incorporating predictive uncertainty into candidate selection, the approach aims to identify reliable regions of chemical space, prioritize informative experiments, and eliminate unpromising candidates before physical synthesis. This integrated computational–experimental strategy establishes a pathway toward the inverse design of electrochromic polymers with targeted spectroelectrochemical properties, shifting materials discovery from trial-and-error exploration toward uncertainty-aware, autonomous, and data-driven design.