The user wants me to rewrite a news headline for a mobile feed based on the given rules.

Short summary

Researchers propose a Context-Augmented Prompting framework where GNN experts provide predictive hints and explanatory subgraphs to small language models for molecular property prediction from SMILES strings. Enriching prompts with graph-derived context yields up to 74% relative accuracy improvement on Tox21, though a gap remains to specialized GNN models. The work highlights both the value and limits of text-conditioned reasoning for molecular structure.

• •SLMs suffer structural blindness when predicting molecular properties from SMILES strings
• •GNN-derived context hints and explanatory subgraphs boost accuracy up to 74% on Tox21
• •Persistent gap to specialized GNN models remains, showing limits of text-conditioned molecular reasoning

Generated with AI, which can make mistakes.