AVTrace - Cat7 Event-conditioned Comprehension - Answer ROUGE-L: leaderboard
AVTrace Cat7: Event-conditioned temporal grounding and comprehension, measured by Answer ROUGE-L (0–1) over 1,000 test questions. Five off-the-shelf omni-model runs use greedy decoding with 512 new tokens; source-native interfaces expose full clips, sampled frames/audio, or only 30 seconds depending on the model. A reference-blind Qwen3.5-27B normalizes answers before deterministic scoring against partly model-assisted silver references. Invalid answers receive zero; exact checkpoint revisions remain unreported.
Metric: Answer ROUGE-L (0–1). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | InteractiveOmni-4B (AVTrace Run: paper v1 off-the-shelf; first 30 seconds; greedy 512 tokens; exact checkpoint unreported) | 0.22 | |
| 2 | Qwen3-Omni-30B (AVTrace Run: paper v1 off-the-shelf; full clips via native interface; greedy 512 tokens; exact checkpoint unreported) | 0.21 | |
| 3 | gemma-4-E4B-it (AVTrace Run: paper v1 off-the-shelf; first 30 seconds; greedy 512 tokens; exact checkpoint unreported) | 0.21 | |
| 4 | MiniCPM-o-4.5 (AVTrace Run: paper v1 off-the-shelf; full clips via native interface; greedy 512 tokens; exact checkpoint unreported) | 0.18 | |
| 5 | Qwen2.5-Omni-3B (AVTrace Run: paper v1 off-the-shelf; up to 300 uniform frames and 300-second audio window; greedy 512 tokens; exact checkpoint unreported) | 0.14 |
Interactive version: theaggregate.ai/benchmark?slug=avtrace-cat7-event-conditioned-comprehension-answer-rouge-l · How It Works · Data refreshed daily, snapshot 2026-11-05.