In materials research, it is often necessary to reason across papers, supplementary information, figures, microstructures, XRD, spectra, and simulation snapshots. For this task, it may be useful to separate image files and metadata in storage, and design an agent that calls the relevant figure only when needed.
The most practical applications are literature review and data interpretation. For example, one could group figures from multiple papers on a specific composition family and have the agent carry out 50~100 turns of questioning while comparing and explaining them. Evaluation can focus on whether the agent does not lose figure references, does not invoke the same figure redundantly, and continues to cite evidence accurately after long dialogue.