MaterIA: Learnable Tool Injection for Emergent Tool Discrimination in Molecular Transformers

This work introduces MaterIA, a lightweight transformer (658K parameters) that predicts molecular properties directly from raw atomic data, atomic numbers, 3D coordinates, and pairwise distances without pre-programmed physical rules or domain-specific inductive biases. The key contribution is Learnable Tool Injection, a framework that offers physics-informed tools (spatial, electronegativity, and angular biases) as optional attention biases, each gated by a learnable scale parameter (bias_scale). The model autonomously determines which tools to use for each target property. On the QM9 benchmark (134K molecules), MaterIA exhibits emergent tool discrimination: spatial tools reduce HOMO-LUMO band gap prediction error by 72.3% while degrading internal energy (U0) predictions by 4.2%; conversely, electronegativity and angular tools improve U0 predictions by 29.9% and 30.7%, respectively. This selective behavior emerges without explicit supervision no loss term, reward signal, or architectural constraint guides tool selection. Analysis of learned bias_scale parameters reveals that the model amplifies useful tools (mean bias_scale of 2.35 for band gap+spatial) while moderating less relevant ones, with amplification magnitude correlating with prediction improvement. These results demonstrate that a transformer can learn not only to predict molecular properties but also to identify which physical information is relevant to each property, suggesting a path toward self-organizing molecular models that autonomously build their own physical intuition.