LLaVA-Video-72B: benchmark results
Provider: LLaVA Team. Access: Open.
Aggregate Rating 1586 ± 34, rank #1163 of 10916 rated models, from 28 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| SFI-Bench | 64.9 | Macro-average accuracy (%) over the six tasks (global and co | 100 |
| SFI-Bench - Operation Planning | 58.4 | Accuracy (%) on the operation planning questions, four-optio | 96.9 |
| Flat-Pack Bench - Part Mating | 56.32 | Mate: decide whether two parts are connected in the final as | 91.4 |
| SFI-Bench - Global and Conditional Counting | 57.9 | Accuracy (%) on the global and conditional counting question | 87.5 |
| Flat-Pack Bench: Evaluating Spatio-Temporal Understanding in Large Vision-Language Models through Furniture Assembly | 37.54 | Micro Avg. (%) | 75.8 |
| SFI-Bench - Cross-View Path Reasoning | 70.3 | Accuracy (%) on the cross-view multi-hop path reasoning ques | 70.8 |
| TOC-Bench - Event Existence | 49.3 | Event Existence accuracy (%) on TOC-Bench (2,323 human-verif | 70 |
| Flat-Pack Bench - Part Tracking | 35.8 | Track: recover the correspondence of shuffled part IDs betwe | 69 |
| Flat-Pack Bench - Temporal Ordering | 36.77 | TOrd: order the connection events of the highlighted parts, | 69 |
| SFI-Bench - Troubleshooting | 50.9 | Accuracy (%) on the causal hypothesis and troubleshooting qu | 62.5 |
| TVBench | 50 | Average Accuracy | 55.6 |
| SFI-Bench - Functional Association | 56.7 | Accuracy (%) on the functional association questions, four-o | 54.2 |
Interactive version: theaggregate.ai/model?slug=llava-video-72b · How It Works · Data refreshed daily, snapshot 2026-11-08.