Tag
Introduces SMMBench, a benchmark to evaluate multimodal agents' ability to retrieve, align, and compose evidence scattered across independently originated sources like conversations, tables, and documents. Experiments show current systems struggle with this source-distributed memory composition task.