video-SALMONN 2: benchmark results
Provider: Tsinghua & ByteDance. Released 2025-07-08. Access: Open.
Aggregate Rating 1479 ± 48, rank #2673 of 10916 rated models, from 29 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| TraceAV-Bench - Visual Counting | 39.38 | Accuracy (%) on the 226 Visual Counting (VC) questions, four | 82.1 |
| SVHalluc - Temporal Forecasting | 33.33 | Accuracy (%; 276 option-balanced video questions, when the n | 80 |
| video-SALMONN-2 Test Set (GPT-4.1 judge) - Total error | 38.8 | Missed + hallucinated events (%) | 75 |
| Daily-Omni (caption QA, Gemini-2.5-Pro reader) | 29.9 | Accuracy (%) | 70 |
| WorldSense (caption QA, Gemini-2.5-Pro reader) | 18.2 | Accuracy (%) | 70 |
| SVHalluc - Fine-Grained Semantic Alignment | 79.58 | Accuracy (%; 444 option-balanced video questions, whether a | 60 |
| UGC-VideoCap - Average | 67.2 | Mean judge score (audio, visual, detail) | 58.3 |
| TraceAV-Bench - Spatiotemporal Localization | 31.28 | Accuracy (%) on the 227 Spatiotemporal Localization (SL) que | 46.4 |
| TraceAV-Bench - Temporal Sequencing | 41.24 | Accuracy (%) on the 97 Temporal Sequencing (TS) questions, f | 42.9 |
| SVHalluc - Cross-Modal Semantic Binding | 72.77 | Accuracy (%; 534 option-balanced video questions, whether an | 40 |
| SVHalluc - Cross-Modal Temporal Binding | 48.58 | Accuracy (%; 339 option-balanced video questions, temporal b | 40 |
| SVHalluc - Global Semantic Alignment | 53.92 | Accuracy (%; 444 option-balanced video questions, whether th | 40 |
Interactive version: theaggregate.ai/model?slug=video-salmonn-2 · How It Works · Data refreshed daily, snapshot 2026-11-08.