Context-Augmented Prompting Boosts Small LM Molecular Accuracy by 74%

A team from NCSR Demokritos (Bougiatiotis, Kelesis, Paliouras) introduces Context-Augmented Prompting, a modular agentic framework that equips small language models (SLMs) with graph-based tools at inference time. The system pairs a trained GNN expert model that outputs a predictive hint with confidence scores, along with a GNN that extracts instance-specific explanatory subgraphs (e.g., a subgraph SMILES and a textual paragraph describing the motif). This allows SLMs that natively suffer from "structural blindness"—they see molecules as flat SMILES strings—to incorporate key topological cues from graph representations.

Evaluated on the MUTAG and Tox21 benchmarks under five prompting configurations, the approach yields substantial accuracy gains. Relative improvements exceed 25% across both datasets, with a striking 74% lift on Tox21. The researchers also validated the functional relevance of extracted motifs through a necessity-based edge-drop intervention, confirming that the identified subgraphs are causally linked to the target properties. Despite these gains, a persistent gap to specialized GNN models remains, underscoring both the value and the limitations of text-conditioned reasoning for molecular structure prediction.

The work was presented at the 2nd Causal Neuro-symbolic AI Workshop (ESWC2026). It offers a practical path for leveraging lightweight SLMs in drug discovery and materials science—without the compute overhead of large models—while also exposing the fundamental trade-off between textual reasoning and graph-native understanding.