UR-FUNNY-Temporal: leaderboard

Laughter localization on UR-FUNNY-Temporal (2026), TED talk clips from UR-FUNNY-V2 with temporal laughter annotations: a zero-shot model is asked whether laughter occurs in a 5-second video clip and, if so, when it starts and ends. Precision at IoU 0.5 on clips with laughter, in percent; higher is better.

Metric: Laughter localization precision at IoU 0.5 (times 100, so 0-100): share of positive test segments whose predicted laughter interval overlaps a ground-truth laughter event with IoU of at least 0.5 (maximum over the segment's events); the model is prompted, zero-shot, to say whether laughter occurs in a 5-second video segment and give its start and end time, on the UR-FUNNY-Temporal test split (TED talk clips of UR-FUNNY-V2); higher is better. Source: arxiv.org. Saturation forecast: Around March 2027. 3 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Gemini 3 Flash39.3#175
2Qwen2.5-Omni-7B21.8#1884
3MTLLFM Qwen2-Audio-7B (checkpoint unspecified)13.7

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=ur-funny-temporal · How It Works · Data refreshed daily, snapshot 2026-11-21.