Preprint recombines model outputs to produce additional verified molecular designs

Researchers report in a preprint that Structured Re-Decoding produced verified molecules absent from a pretrained model's original sample pool. The method requires no additional training: it builds local candidates from model scores, samples or external specifications, jointly selects them and verifies the reconstructed outputs. The researchers report improved structured inference across molecular generation, protein design, RNA structure prediction and antibody modeling. External chemical templates also support noncanonical amino acids, peptide cyclization and molecular construction under grammar rules. Experiments using a physical coherent Ising machine to perform optimization remain preliminary.