Exposure bias is the mismatch between training and generation in sequence models: training uses ground-truth histories, while inference must continue from the model’s own predictions, including errors. In molecular generation, this affects decoders that reconstruct SMILES token by token. section 10.4.3
A teacher-forced molecular decoder follows this pattern:
- A mistaken token can therefore place later predictions in histories the decoder rarely encountered during training. section 10.4.3
