Streaming Video Understanding on RVS-Movie
61.6AccuracyReKV
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| ReKVBackbone=Qwen2.5-VL-7B, Encoding=Full, Retrieval=Internal, Video Enc.=8.47 FPS, Latency=2.8s, GPU=29 GB, KV-Cache=11.1 GB/h2026.02 | 61.6 | 3.65 | |
| MemStreamBackbone=Qwen2.5-VL-7B, Encoding=AKS, Retrieval=MoE, Video Enc.=8.68 FPS, Latency=2.6s, GPU=32 GB, KV-Cache=11.1 GB/h2026.02 | 59.7 | 3.6 | |
| MemStreamBackbone=Qwen2.5-VL-7B, Encoding=AKS, Retrieval=Internal, Video Enc.=8.50 FPS, Latency=2.6s, GPU=29 GB, KV-Cache=11.1 GB/h2026.02 | 59.1 | 3.6 | |
| ReKVBase Model=LLaVA-OV-7B, Upper Bound=True2026.01 | 54.4 | 3.6 | |
| HERMESBase Model=LLaVA-OV-7B, Token Budget=6K tokens2026.01 | 54.4 | 3.6 | |
| HERMESBase Model=LLaVA-OV-7B, Token Budget=4K tokens2026.01 | 54.4 | 3.6 | |
| ReKVCPU offloading=true2025.10 | 54.4 | 3.6 | |
| StreamingTOMBase Model=LLaVA-OV-7B2026.01 | 53.2 | 3.5 | |
| StreamingTOM2025.10 | 53.2 | 3.5 | |
| Flash-VStreamBase Model=LLaVA-OV-7B2026.01 | 53.1 | 3.3 | |
| Flash-VStream2025.10 | 53.1 | 3.3 | |
| StreamMemBase Model=LLaVA-OV-7B2026.01 | 52.7 | 3.4 | |
| StreamMem2025.10 | 52.7 | 3.4 | |
| InfiniPot-VBase Model=LLaVA-OV-7B2026.01 | 51.4 | 3.5 | |
| InfiniPot-V2025.10 | 51.4 | 3.5 | |
| ReKV w/o off.Base Model=LLaVA-OV-7B2026.01 | 50.8 | 3.4 | |
| ReKVCPU offloading=false2025.10 | 50.8 | 3.4 | |
| ReKVBase Model=LLaVA-OV-0.5B, Upper Bound=True2026.01 | 44.6 | 3.4 | |
| LLaVA-OV-7BBase Model=LLaVA-OV-7B2026.01 | 43 | 3.3 | |
| HERMESBase Model=LLaVA-OV-0.5B, Token Budget=6K tokens2026.01 | 42.5 | 3.4 | |
| HERMESBase Model=LLaVA-OV-0.5B, Token Budget=4K tokens2026.01 | 41.7 | 3.4 | |
| LLaVA-OV-0.5BBase Model=LLaVA-OV-0.5B2026.01 | 37.2 | 3.2 | |
| Qwen2.5-VL-7BMode=Traditional Streaming Prefill (Baseline), acn=∞, ur=∞, s=–2026.06 | 34.6 | — | |
| Qwen2.5-VL-3BMode=Traditional Streaming Prefill (Baseline), acn=∞, ur=∞, s=–2026.06 | 33.91 | — | |
| ViCoStreamBackbone=Qwen2.5-VL-7B, acn=8, ur=64, s=0.25 (↓48.4%)2026.06 | 33.23 | — | |
| ViCoStreamBackbone=Qwen2.5-VL-7B, acn=8, ur=32, s=0.25 (↓48.4%)2026.06 | 33.02 | — | |
| ViCoStreamBackbone=Qwen2.5-VL-3B, acn=8, ur=64, s=0.1 (↓73.0%)2026.06 | 32.25 | — | |
| ViCoStreamBackbone=Qwen2.5-VL-3B, acn=4, ur=32, s=0.25 (↓57.6%)2026.06 | 31.98 | — | |
| ViCoStreamBackbone=Qwen2.5-VL-3B, acn=4, ur=64, s=0.25 (↓57.6%)2026.06 | 31.97 | — | |
| ViCoStreamBackbone=Qwen2.5-VL-7B, acn=8, ur=32, s=0.1 (↓70.5%)2026.06 | 31.76 | — | |
| ViCoStreamBackbone=Qwen2.5-VL-7B, acn=4, ur=32, s=0.25 (↓48.0%)2026.06 | 31.7 | — | |
| ViCoStreamBackbone=Qwen2.5-VL-7B, acn=4, ur=64, s=0.25 (↓48.0%)2026.06 | 31.67 | — | |
| ViCoStreamBackbone=Qwen2.5-VL-3B, acn=8, ur=64, s=0.25 (↓57.3%)2026.06 | 31.55 | — | |
| ViCoStreamBackbone=Qwen2.5-VL-3B, acn=8, ur=32, s=0.1 (↓72.9%)2026.06 | 31.34 | — | |
| ViCoStreamBackbone=Qwen2.5-VL-7B, acn=8, ur=64, s=0.1 (↓70.5%)2026.06 | 31.07 | — | |
| ViCoStreamBackbone=Qwen2.5-VL-3B, acn=8, ur=32, s=0.25 (↓57.3%)2026.06 | 30.93 | — | |
| ViCoStreamBackbone=Qwen2.5-VL-3B, acn=4, ur=64, s=0.1 (↓72.9%)2026.06 | 29.04 | — | |
| ViCoStreamBackbone=Qwen2.5-VL-3B, acn=4, ur=32, s=0.1 (↓72.9%)2026.06 | 28.39 | — | |
| ViCoStreamBackbone=Qwen2.5-VL-7B, acn=4, ur=32, s=0.1 (↓70.6%)2026.06 | 28.3 | — | |
| ViCoStreamBackbone=Qwen2.5-VL-7B, acn=4, ur=64, s=0.1 (↓70.6%)2026.06 | 28.14 | — |