M3Exam
M3Exam: Benchmarking Multimodal Memory for Realistic User-Agent Interactions
Zhengjun Huang, Wenxuan Liu, Zhoujin Tian, Wei Chen, Junle Chen, Yuqian Wu, et al. · 2026-06-05
- Modality
- Image
- Data
- Hybrid (real + synthetic/simulated)
- Task · ground truth
- QA/Ret. · Ans.+Ev.
- Scope
- Memory-specific benchmark
What it tests
- CECross-modal evidenceDirectly evaluated
- LELong-range evidenceDirectly evaluated
- SCState changesPartially evaluated
- CFConflicting memoriesNot evaluated
- UQUnanswerable queriesNot evaluated
- MAMemory-grounded actionNot evaluated
- FGFine-grained visual detailDirectly evaluated
- ETEntity trackingDirectly evaluated
- VCVisual state changePartially evaluated
Survey notes
Multi-session user-agent histories include images and documents; questions have supporting-fact rounds and test memory retrieval, cross-session reasoning and file management.