Tag
Introduces Q-Steer, a rollout-time action-value guidance for molecular policy optimization that uses a frozen prefix-action value scorer to improve sampling without changing the online oracle budget.
This paper systematically compares BPE and Unigram-LM tokenization methods for SMILES strings in chemistry, finding they produce near-disjoint vocabularies and differ significantly in segmentation granularity. The results show that the choice of subword algorithm is a critical modeling decision, not a free default.