Video Understanding on LVB
66.49AccuracyFLoC
Evaluation Results
| Method | Links | |
|---|---|---|
| FLoCModel=Qwen2.5-VL-32B, Max Frames=72002025.10 | 66.49 | |
| TS-LLaVAModel=Qwen2.5-VL-32B, Max Frames=72002025.10 | 65 | |
| LongVUModel=Qwen2.5-VL-32B, Max Frames=72002025.10 | 64.62 | |
| RandomModel=Qwen2.5-VL-32B, Max Frames=72002025.10 | 64.62 | |
| DivPruneModel=Qwen2.5-VL-32B, Max Frames=72002025.10 | 64.32 | |
| DyCokeModel=Qwen2.5-VL-32B, Max Frames=72002025.10 | 63.87 | |
| PruneVidModel=Qwen2.5-VL-32B, Max Frames=72002025.10 | 63.5 | |
| DyCokeModel=Qwen2.5-VL-7B, Max Frames=72002025.10 | 62.98 | |
| LongVUModel=Qwen2.5-VL-7B, Max Frames=72002025.10 | 62.75 | |
| -Model=Qwen2.5-VL-32B, Max Frames=7682025.10 | 62.6 | |
| FLoCModel=Qwen2.5-VL-7B, Max Frames=72002025.10 | 62.6 | |
| PruneVidModel=Qwen2.5-VL-7B, Max Frames=72002025.10 | 62.45 | |
| DivPruneModel=Qwen2.5-VL-7B, Max Frames=72002025.10 | 62.3 | |
| TS-LLaVAModel=Qwen2.5-VL-7B, Max Frames=72002025.10 | 62.08 | |
| RandomModel=Qwen2.5-VL-7B, Max Frames=72002025.10 | 61.63 | |
| -Model=Qwen2.5-VL-7B, Max Frames=7682025.10 | 60.51 | |
| FLoCModel=Qwen2.5-VL-7B, Compression Ratio=2^-32025.10 | 58.12 | |
| DiEPExpert skipping ratio (ρ)=0.6, Number of experts (k)=3, Backbone=InternVL-3.5-30B-A3B-HF2025.11 | 57.84 | |
| Top-k RoutingExpert skipping ratio (ρ)=0, Number of experts (k)=8, Backbone=InternVL-3.5-30B-A3B-HF2025.11 | 57.64 | |
| FastVIDModel=Qwen2.5-VL-7B, Compression Ratio=2^-32025.10 | 57.14 | |
| DivPruneModel=InternVL3-8B, Compression Ratio=2^-32025.10 | 56.92 | |
| FLoCModel=InternVL3-8B, Compression Ratio=2^-32025.10 | 56.69 | |
| FrameOracleFrame=64 → 15.6, Backbone=LLaVA-OneVision2025.10 | 56.5 | |
| FLoCModel=InternVL3-8B, Compression Ratio=2^-42025.10 | 56.47 | |
| DivPruneModel=Qwen2.5-VL-7B, Compression Ratio=2^-32025.10 | 56.17 | |
| MoDESExpert skipping ratio (ρ)=0.6, Number of experts (k)=3, Backbone=InternVL-3.5-30B-A3B-HF2025.11 | 56.12 | |
| MC-MoEExpert skipping ratio (ρ)=0.6, Number of experts (k)=3, Backbone=InternVL-3.5-30B-A3B-HF2025.11 | 56.02 | |
| TS-LLaVAModel=InternVL3-8B, Compression Ratio=2^-32025.10 | 56.02 | |
| RandomModel=InternVL3-8B, Compression Ratio=2^-32025.10 | 56.02 | |
| MoDESExpert skipping ratio (ρ)=0.73, Number of experts (k)=2, Backbone=InternVL-3.5-30B-A3B-HF2025.11 | 55.85 | |
| Top-k RoutingExpert skipping ratio (ρ)=0.6, Number of experts (k)=3, Backbone=InternVL-3.5-30B-A3B-HF2025.11 | 55.81 | |
| NAEEExpert skipping ratio (ρ)=0.6, Number of experts (k)=3, Backbone=InternVL-3.5-30B-A3B-HF2025.11 | 55.74 | |
| RandomModel=Qwen2.5-VL-7B, Compression Ratio=2^-32025.10 | 55.72 | |
| DyCokeModel=InternVL3-8B, Compression Ratio=2^-32025.10 | 55.72 | |
| LongVUModel=Qwen2.5-VL-7B, Compression Ratio=2^-32025.10 | 55.42 | |
| LongVUModel=InternVL3-8B, Compression Ratio=2^-32025.10 | 55.35 | |
| FLoCModel=Qwen2.5-VL-7B, Compression Ratio=2^-42025.10 | 55.27 | |
| TS-LLaVAModel=Qwen2.5-VL-7B, Compression Ratio=2^-32025.10 | 55.2 | |
| DyCokeModel=Qwen2.5-VL-7B, Compression Ratio=2^-32025.10 | 55.12 | |
| ScissorModel=InternVL3-8B, Compression Ratio=2^-32025.10 | 55.12 | |
| STTMModel=InternVL3-8B, Compression Ratio=2^-32025.10 | 54.9 | |
| DivPruneModel=InternVL3-8B, Compression Ratio=2^-42025.10 | 54.82 | |
| STTMModel=Qwen2.5-VL-7B, Compression Ratio=2^-32025.10 | 54.6 | |
| MoDESExpert skipping ratio (ρ)=0.85, Number of experts (k)=1, Backbone=InternVL-3.5-30B-A3B-HF2025.11 | 54.49 | |
| FastVIDModel=Qwen2.5-VL-7B, Compression Ratio=2^-42025.10 | 54.23 | |
| RandomModel=InternVL3-8B, Compression Ratio=2^-42025.10 | 54.23 | |
| FLoCModel=InternVL3-8B, Compression Ratio=2^-52025.10 | 54.23 | |
| DiEPExpert skipping ratio (ρ)=0.73, Number of experts (k)=2, Backbone=InternVL-3.5-30B-A3B-HF2025.11 | 54.22 | |
| PruneVidModel=Qwen2.5-VL-7B, Compression Ratio=2^-32025.10 | 54.15 | |
| ScissorModel=Qwen2.5-VL-7B, Compression Ratio=2^-32025.10 | 54.08 | |
| DivPruneModel=Qwen2.5-VL-7B, Compression Ratio=2^-42025.10 | 54 | |
| TS-LLaVAModel=InternVL3-8B, Compression Ratio=2^-42025.10 | 53.85 | |
| DyCokeModel=Qwen2.5-VL-7B, Compression Ratio=2^-42025.10 | 53.78 | |
| ScissorModel=InternVL3-8B, Compression Ratio=2^-42025.10 | 53.77 | |
| RandomModel=Qwen2.5-VL-7B, Compression Ratio=2^-42025.10 | 53.63 | |
| ScissorModel=Qwen2.5-VL-7B, Compression Ratio=2^-42025.10 | 53.55 | |
| PruneVidModel=InternVL3-8B, Compression Ratio=2^-32025.10 | 53.48 | |
| TS-LLaVAModel=InternVL3-8B, Compression Ratio=2^-52025.10 | 53.33 | |
| DyCokeModel=InternVL3-8B, Compression Ratio=2^-42025.10 | 53.1 | |
| FLoCModel=Qwen2.5-VL-7B, Compression Ratio=2^-52025.10 | 53.1 | |
| MC-MoEExpert skipping ratio (ρ)=0.73, Number of experts (k)=2, Backbone=InternVL-3.5-30B-A3B-HF2025.11 | 53.08 | |
| FastVIDModel=Qwen2.5-VL-7B, Compression Ratio=2^-52025.10 | 52.95 | |
| DivPruneModel=InternVL3-8B, Compression Ratio=2^-52025.10 | 52.88 | |
| NAEEExpert skipping ratio (ρ)=0.73, Number of experts (k)=2, Backbone=InternVL-3.5-30B-A3B-HF2025.11 | 52.79 | |
| LongVUModel=Qwen2.5-VL-7B, Compression Ratio=2^-42025.10 | 52.73 | |
| TS-LLaVAModel=Qwen2.5-VL-7B, Compression Ratio=2^-42025.10 | 52.51 | |
| STTMModel=InternVL3-8B, Compression Ratio=2^-52025.10 | 52.43 | |
| PruneVidModel=InternVL3-8B, Compression Ratio=2^-42025.10 | 52.28 | |
| STTMModel=InternVL3-8B, Compression Ratio=2^-42025.10 | 52.28 | |
| DivPruneModel=Qwen2.5-VL-7B, Compression Ratio=2^-52025.10 | 52.28 | |
| RandomModel=InternVL3-8B, Compression Ratio=2^-52025.10 | 52.13 | |
| Final 10K DPO Recipe (Ours)Recipe=Combined CTP, FV-D, and FV-A-L2026.05 | 52.1 | |
| DyCokeModel=InternVL3-8B, Compression Ratio=2^-52025.10 | 51.98 | |
| PruneVidModel=Qwen2.5-VL-7B, Compression Ratio=2^-42025.10 | 51.83 | |
| STTMModel=Qwen2.5-VL-7B, Compression Ratio=2^-42025.10 | 51.68 | |
| DPO w/ SP + FV-DRecipe=DPO with SFT-policy negatives and video preference data2026.05 | 51.5 | |
| ScissorModel=Qwen2.5-VL-7B, Compression Ratio=2^-52025.10 | 51.46 | |
| DPO w/ CTP + FV-DRecipe=DPO with counterfactual and video preference data2026.05 | 51.4 | |
| DyCokeModel=Qwen2.5-VL-7B, Compression Ratio=2^-52025.10 | 51.38 | |
| LongVUModel=InternVL3-8B, Compression Ratio=2^-42025.10 | 51.31 | |
| ScissorModel=InternVL3-8B, Compression Ratio=2^-52025.10 | 51.31 | |
| DPO w/ CTP + FV-D + LV-MCQARecipe=DPO with counterfactual, video preference, and MCQA data2026.05 | 51.1 | |
| LongVUModel=InternVL3-8B, Compression Ratio=2^-52025.10 | 51.01 | |
| DPO w/ SPRecipe=DPO with SFT-policy negatives2026.05 | 50.9 | |
| DPO w/ CTP + FV-D + FV-ARecipe=DPO with counterfactual, video preference, and audio preference data2026.05 | 50.8 | |
| TS-LLaVAModel=Qwen2.5-VL-7B, Compression Ratio=2^-52025.10 | 50.49 | |
| DPO w/ OP + FV-D + LV-MCQARecipe=DPO with original-sync, video preference, and MCQA data2026.05 | 50.4 | |
| LongVUModel=Qwen2.5-VL-7B, Compression Ratio=2^-52025.10 | 50.34 | |
| STTMModel=Qwen2.5-VL-7B, Compression Ratio=2^-52025.10 | 50.11 | |
| PruneVidModel=InternVL3-8B, Compression Ratio=2^-52025.10 | 49.96 | |
| RandomModel=Qwen2.5-VL-7B, Compression Ratio=2^-52025.10 | 49.89 | |
| PruneVidModel=Qwen2.5-VL-7B, Compression Ratio=2^-52025.10 | 49.66 | |
| Qwen3-Omni-30BRecipe=Vanilla Baseline2026.05 | 49.1 | |
| Top-k RoutingExpert skipping ratio (ρ)=0.73, Number of experts (k)=2, Backbone=InternVL-3.5-30B-A3B-HF2025.11 | 48.87 | |
| MC-MoEExpert skipping ratio (ρ)=0.85, Number of experts (k)=1, Backbone=InternVL-3.5-30B-A3B-HF2025.11 | 48.31 | |
| DPO w/ OP + SPRecipe=DPO with original-sync and SFT-policy negatives2026.05 | 47.7 | |
| NAEEExpert skipping ratio (ρ)=0.85, Number of experts (k)=1, Backbone=InternVL-3.5-30B-A3B-HF2025.11 | 47.42 | |
| DiEPExpert skipping ratio (ρ)=0.85, Number of experts (k)=1, Backbone=InternVL-3.5-30B-A3B-HF2025.11 | 47.26 | |
| MovieChatFrame=512 → 64, Backbone=LLaVA-OneVision2025.10 | 44.2 | |
| SFT w/ CTP + FV-D + FV-ALRecipe=SFT with counterfactual, general video preference, and audio-visual data2026.05 | 40.8 |