An autoregressive model generates a sequence one token at a time. Each prediction conditions on the tokens already produced. For molecular generation, a unidirectional decoder builds a SMILES string from left to right until it emits an end-of-sequence token. section 10.4.3
A VAE-based molecular generation workflow uses autoregression as its decoding stage:
- Tokenize a SMILES string with SmilesPE and encode it with a bidirectional GRU. section 10.4.5
- Sample a latent vector z from the encoder's latent distribution using the reparameterization trick. section 10.4.5
- Evaluate generated molecules with validity, uniqueness, novelty, and diversity metrics. section 10.4.5
Training can use teacher forcing: the decoder receives ground-truth tokens rather than its own previous predictions. This improves stability and convergence but creates exposure bias, because inference supplies no ground truth. Word dropout exposes the model to imperfect inputs and improves recovery from errors. section 10.4.3
At inference, greedy decoding selects the most probable token and is fast but can produce suboptimal sequences. Temperature sampling trades conservatism for diversity: lower values narrow choices, while higher values broaden them. section 10.4.3







