APM-Bench: 549-session benchmark exposes utility-latency-storage trade-off in streaming video memory
Jianguo Huang · hf · 2026-09-30
APM-Bench reformulates real-world streaming interaction as multi-session life trajectories to test persistent memory for egocentric video assistants: 549 sessions, 104 trajectories, and 2,719 candidates spanning objective and open-ended questions. Evaluation of general video models and specialized streaming memory systems reveals a clear utility-latency-storage trade-off — no existing method achieves reliable long-term recall, low overhead, and effective proactive assistance simultaneously, and models still struggle to acknowledge missing evidence.
More from Research
- Image Lab Tops Robot Leaderboard: FLUX 3 Action Runs on a Desk-Size Arm — lmoroney · 2026-09-30
- MIT builds AI-powered Raman 'barcode' to noninvasively detect senescent 'zombie cells' — MacrinePhD · 2026-09-30
- Anil Seth's biological naturalism essay against AI consciousness draws mass commentary collection — anilkseth · 2026-09-30
- Causal Inference Assumptions: A Practical Checklist Before You Run the Model — Nasereliver · 2026-09-30
- Sparse-View Gaussian Splatting Demo Shows 3D Reconstruction From Few Images — enndeeee · 2026-09-30
- Open-source self-driving guide adds VFH: binning LiDAR into 5° obstacle histograms — 4310sy · 2026-09-30