Abstract
<jats:p>Deep learning has transformed chemistry, yet most chemical models remain narrow and task-specific, requiring full retraining for every new target such as molecular property regression, reaction-outcome prediction, or membrane-design ranking. Meanwhile, large vision-language models have shown a remarkable ability to solve novel tasks at inference time through in-context learning, adapting to a handful of demonstrations without any parameter update. In this work we propose ChemICL, a multimodal in-context learner that transplants this paradigm into chemistry. ChemICL encodes molecular, visual, and tabular modalities into a shared latent space and uses a confidence-guided retrieval module to select the most informative in-context demonstrations for each query, after which a frozen backbone with a lightweight task head produces the prediction. No task-specific fine-tuning is performed at inference. Training relies on a maximum-normalized element-wise ratio objective that stabilizes large-batch optimization across heterogeneous modalities. We evaluate ChemICL on three chemistry tasks: molecular property prediction on ESOL solubility and graphene-membrane permeability, photocatalytic reaction-outcome prediction, and membrane-design ranking. Across all tasks ChemICL consistently outperforms strong baselines including chemprop, a fine-tuned SciBERT, and a vision-language zero-shot model, and its performance improves monotonically as more in-context exemplars are provided. Ablation studies confirm the contribution of each modality and the benefit of confidenceguided demonstration filtering, while cross-task experiments show that chemical structure learned in one task transfers to others. These results establish multimodal in-context learning as a viable and data-efficient paradigm for chemistry discovery.</jats:p>