MolReps: Teaching LLMs to Convert Between Molecular String Representations
Researchers introduce MolReps, a pipeline of 9,600 question-answer pairs spanning ten molecular representations for tasks like conversion, normalization, and 2D structure recognition. Post-training Qwen3-4B-based text-only and multimodal models lifted InChIKey match accuracy on a molecule-overlap test set from 4.01% to 75.13% and from 7.11% to 73.14%, respectively. However, on a molecule-disjoint test set InChIKey match stayed low or even dropped, even as output validity roughly tripled, and the models improved about 27.84% relative on average over the base model on MolRepBench for property prediction.
Quantifies both the gains and the generalization limits of targeted post-training that teaches LLMs to handle equivalent molecular encodings, a concrete advance for AI-for-chemistry readers.
Comments & Questions (0)
No comments yet
Be the first to comment!