SignalTech/AI
Emergent Biological Capabilities in a Foundation Model for Molecular Interactions
High-priority alert based on source significance and keyword relevance.
Context
What happened: Researchers introduced Bonbon, a foundation model for molecular interactions trained on about 3.2 trillion tokens of paired protein-ligand sequences, accepted to the NeurIPS 2026 main track as an Oral.
Key numbers: 3.2 trillion tokens; six model scales from 42M to 1.6B parameters; 1:2,000 parameter-to-token ratio; 61.5% dose-response
Why it matters: The authors report that every capability shows a sharp jump at the largest scale even as pretraining loss decreases smoothly, which they say shows loss-based scaling laws alone cannot predict when zero-shot capabilities appear.
What changed
Researchers introduced Bonbon, a foundation model for molecular interactions trained on about 3.2 trillion tokens of paired protein-ligand sequences, accepted to the NeurIPS 2026 main track as an Oral.
Why it matters
The authors report that every capability shows a sharp jump at the largest scale even as pretraining loss decreases smoothly, which they say shows loss-based scaling laws alone cannot predict when zero-shot capabilities appear.