Tag
This paper investigates whether tabular in-context learning models, pretrained on synthetic causal tables, can generalize to predict biomolecular properties from limited labeled data. The authors find that these models are competitive for protein fitness regression but that representation choice is crucial for small-molecule classification.