DMV-Bench
DMV-Bench: Diagnosing Long-Horizon Multimodal Agents' Visual Memory with Incidental Cue Injection
Yujin Tang, Chenming Shang, Ruize Xu, Nikhil Singh · 2026-06-25
- Modality
- GUI & computer use
- Data
- Simulated
- Task · ground truth
- Ret./Act. · State+Act.+Prov.
- Scope
- Memory-specific benchmark
What it tests
- CECross-modal evidencePartially evaluated
- LELong-range evidenceDirectly evaluated
- SCState changesPartially evaluated
- CFConflicting memoriesNot evaluated
- UQUnanswerable queriesNot evaluated
- MAMemory-grounded actionDirectly evaluated
- GSPersistent interface stateDirectly evaluated
- AOAction–outcome associationPartially evaluated
- XRCross-session reuseDirectly evaluated
Survey notes
Incidental image cues observed in earlier shopping sessions must be recalled to navigate to the exact product; cues are excluded from text channels.