Eight Tokens for One Group

Motif-Merged SMILES Vocabularies Shift Which Fluorinated Structures a Generative Model Proposes

A research poster from the School of Emergent Priorities trains two otherwise identical molecular generators on 1.4 million structures, differing only in whether six fluorinated motifs hold their own token or are spelled out character by character. Merging shortens fluorine-bearing strings by 13% and lifts validity from 94.1% to 95.3%, but the sampled distribution follows the vocabulary rather than the corpus: trifluoromethyl appears in 31.6% of samples against 24.1% of training molecules, while aryl monofluorine, left spelled out, falls from 37.2% to 29.8%.

@misc{slop_gz6viw,
  author       = {Sunniva Balogun and Ylva Onyekwere and Eskil Nwabueze},
  title        = {Eight Tokens for One Group: Motif-Merged SMILES Vocabularies Shift Which Fluorinated Structures a Generative Model Proposes},
  year         = {2026},
  publisher    = {Slop University},
  doi          = {10.5555/slop.gz6viw},
  url          = {https://slop.university/outputs/slop-poster-a-chemical-language-gz6viw/},
  version      = {1.0},
  note         = {Research poster},
}