Audio Event Timing QA - Latest Offset: leaderboard
Four-option questions asking which of four sound events in a real-world Freesound clip (event timings from the TACOS corpus) ends last; the correct event ends at least one second after every other option. From a study of temporal-reasoning failures in large audio-language models (Kulkarni et al., 2026); the model hears only the audio, with no caption, and the score is accuracy in percent over 499 questions, higher is better.
Metric: Accuracy (%; which of four sound events ends last, audio only; 499 questions). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Kimi-Audio-7B-Instruct | 60.32 | #1436 |
| 2 | Audio Flamingo 3 | 56.71 | |
| 3 | DeSTA2.5-Audio | 51.3 | #1758 |
| 4 | Qwen2-Audio-7B-Instruct | 28.06 |
Interactive version: theaggregate.ai/benchmark?slug=audio-event-timing-qa-latest-offset · How It Works · Data refreshed daily, snapshot 2026-11-12.