Infrared (IR) spectroscopy provides fast and non-destructive molecular analysis, but inferring full structures directly from spectra remains highly underdetermined, as distinct molecules often share similar vibrational patterns. Recent Transformer models show promise for IR-to-SMILES prediction, yet still face this intrinsic ambiguity. We address this issue by introducing latent molecular evidence: intermediate symbolic representations that encode mid-level structural cues. Auxiliary predictors infer molecular fingerprints and extended functional group descriptors from spectra and molecular formulas, and a SMILES generator conditions its decoding on these predicted signals. This two-stage multitask design encourages interpretable, data-driven reasoning between raw spectral features and molecular structure. Experiments show that integrating such complementary cues and training the generator under realistically noisy evidence, substantially improves accuracy, robustness, and chemical validity over a strong Transformer baseline.