APM-Bench: Benchmarking cross-Session Persistent Memory for Real-World Egocentric Streaming Video Assistants
Paper · Project page · Code
APM-Bench groups egocentric video into activity-related trajectories that span separate interactions. Its tasks test whether an assistant can use earlier experience while answering questions, perceiving the current scene, and deciding when to respond.
| Video source | Trajectories | Sessions | Candidates |
|---|---|---|---|
| EgoLife | 76 | 472 | 1,896 |
| HD-EPIC | 28 | 77 | 823 |
| Total | 104 | 549 | 2,719 |
| Capability | Candidates |
|---|---|
| Cross-session understanding | 887 |
| Real-time perception | 716 |
| Adaptive response | 1,116 |
The benchmark covers 12 tasks and 5,371 evaluation probes. The counts above describe the complete benchmark; this repository currently contains its 549 session videos.
APM-Bench: Benchmarking cross-Session Persistent Memory for Real-World Egocentric Streaming Video Assistants
Paper · Project page · Code
APM-Bench groups egocentric video into activity-related trajectories that span separate interactions. Its tasks test whether an assistant can use earlier experience while answering questions, perceiving the current scene, and deciding when to respond.
| Video source | Trajectories | Sessions | Candidates |
|---|---|---|---|
| EgoLife | 76 | 472 | 1,896 |
| HD-EPIC | 28 | 77 | 823 |
| Total | 104 | 549 | 2,719 |
| Capability | Candidates |
|---|---|
| Cross-session understanding | 887 |
| Real-time perception | 716 |
| Adaptive response | 1,116 |
The benchmark covers 12 tasks and 5,371 evaluation probes. The counts above describe the complete benchmark; this repository currently contains its 549 session videos.
