Proactive VideoQA on SPOT-Bench
49.5Detection ABD Timeliness Score@5Qwen2.5-VL
Evaluation Results
| Method | Links | ||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Qwen2.5-VL# Frames=16, Post Train?=false, Model Size=7B, Backbone=Qwen2.5-VL [8]2026.04 | 49.5 | 8.3 | 49.4 | 5.4 | 27.5 | 2.9 | 4 | 1.2 | 0 | 0 | 0 | 0 | 21.7 | 3 | |
| Qwen2.5-VL# Frames=4, Post Train?=false, Model Size=7B, Backbone=Qwen2.5-VL [8]2026.04 | 49.3 | 10 | 44.7 | 6.2 | 16.9 | 3.6 | 3.2 | 1.2 | 0 | 0 | 0 | 0 | 19 | 3.5 | |
| Qwen2.5-VL# Frames=1, Post Train?=false, Model Size=7B, Backbone=Qwen2.5-VL [8]2026.04 | 44.9 | 12 | 31.3 | 7.4 | 3.6 | 1.5 | 2.2 | 0.8 | 0 | 0 | 0 | 0 | 13.7 | 3.6 | |
| AsynKV# Frames=1fps, Post Train?=false, Model Size=7B, Backbone=Qwen2.5-VL [8]2026.04 | 40.2 | 22.6 | 36.8 | 19.5 | 30.8 | 18 | 1.2 | 1 | 0 | 0 | 0 | 0 | 18.2 | 10.2 | |
| StreamingVLM# Frames=2fps, Post Train?=true, Model Size=7B, Backbone=Qwen2.5-VL [8]2026.04 | 8.2 | 4.8 | 5.5 | 2.1 | 31.6 | 0.9 | 1.2 | 0.3 | 3.2 | 0.7 | 2.4 | 0.5 | 8.7 | 1.6 | |
| MMDuet2 + KVflush# Frames=1fps, Post Train?=true, Model Size=3B, Backbone=Qwen2.5-VL [8]2026.04 | 6.2 | 7.5 | 3.5 | 4.8 | 37.8 | 5.3 | 2.6 | 3 | 0 | 0 | 0 | 0 | 8.4 | 3.4 | |
| MMDuet2 + StreamingVLM# Frames=1fps, Post Train?=true, Model Size=3B, Backbone=Qwen2.5-VL [8]2026.04 | 1 | 1.7 | 4.9 | 6.6 | 31 | 4 | 1 | 2.2 | 0 | 0 | 0 | 0 | 6.3 | 2.4 | |
| Qwen2.5-VL# Frames=0, Post Train?=false, Model Size=7B, Backbone=Qwen2.5-VL [8]2026.04 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 |