DunphyBench
Long-Horizon Embodied Decision-Making via Multimodal Memory Compression
Bingxuan Li, Rui Yang, Cheng Qian, Jiateng Liu, Jeonghwan Kim, Zhenhailong Wang, et al. · 2026-08-02
- Modality
- Embodied & 3D
- Data
- Simulated
- Task · ground truth
- Plan/Act. · Ev.+Act.
- Scope
- memory dependent task
What it tests
- CECross-modal evidenceDirectly evaluated
- LELong-range evidenceDirectly evaluated
- SCState changesPartially evaluated
- CFConflicting memoriesNot evaluated
- UQUnanswerable queriesNot evaluated
- MAMemory-grounded actionDirectly evaluated
- SMSpatial memoryDirectly evaluated
- OSObject-state trackingPartially evaluated
- TSTask-state trackingDirectly evaluated
Survey notes
Agents must accumulate and compare visual observations across housing environments; evaluation contrasts no-memory, text-memory and multimodal-memory configurations and varies history budget.