MMA-Bench

MMA: Multimodal Memory Agent

Yihao Lu, Wanru Cheng, Zeyu Zhang, Hao Tang · 2026-02-18

Read the paper

Modality

Image

Data

Synthetic

Task · ground truth

QA/Ref. · Ans.+Rel.

Scope

Memory-specific benchmark

What it tests

• CECross-modal evidenceDirectly evaluated
• LELong-range evidenceDirectly evaluated
• SCState changesDirectly evaluated
• CFConflicting memoriesDirectly evaluated
• UQUnanswerable queriesDirectly evaluated
• MAMemory-grounded actionNot evaluated
• FGFine-grained visual detailPartially evaluated
• ETEntity trackingNot evaluated
• VCVisual state changeDirectly evaluated

Survey notes

Ten-session multimodal histories manipulate source reliability, conflicting text/image evidence and resolution, then probe belief revision and justified abstention.