Do Language Models Dream of Binding Molecules? Benchmarking LLMs under Spatial Constraints

This paper evaluates whether general-purpose language models can reason over complex 3D constraints in protein-pocket-conditioned ligand generation. It introduces 3D-Fit, a token-efficient benchmark covering multiple conditioning signals, including anchor fragments, pharmacophore points, and mandatory pocket-ligand interactions. The reported results show that LLM-based methods remain behind state-of-the-art specialized diffusion models, but can handle several spatial constraints simultaneously and may scale to heterogeneous molecular-design settings.