SMMBench: A Benchmark for Source-Distributed Multimodal Agent Memory

Researchers introduced SMMBench, a benchmark for source-distributed multimodal agent memory.

The dataset includes 1,877 samples drawn from 264 sources, with fragmented evidence across sources.

It evaluates cross-source multimodal reasoning, conflict resolution, preference reasoning, and memory-grounded action prediction.

The work exposes a major evaluation gap: current multimodal systems are often tested on curated contexts rather than real-world distributed evidence, and they still perform poorly on these harder tasks.