VoiceGiraffe (Cascaded Captions) - Multi-Hop Reasoning: leaderboard
VoiceGiraffe (2026) tests hour-scale audio understanding with 1,500 four-option questions on long English and Chinese recordings. Under the cascaded-caption protocol the model captions the audio in 30-second windows and then answers from its own concatenated captions; this board is accuracy in percent on the 500 multi-hop questions that link two to four distant segments, higher is better.
Metric: Accuracy (%) on the 500 multi-hop questions (causal alignment and event tracking across two to four distant segments) of VoiceGiraffe on 123 English and Chinese recordings of 30 minutes to 2 hours; cascaded caption aggregation: the model captions the recording in 30-second sliding windows (speech, speaker traits, sound events, music) and then answers from the concatenated captions and the question itself, without an external reasoning model; answers are matched by regular expression; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 12 models tracked. Updated 7 Oct 2026.
Top models
| # | Model | Setting | Score | Overall rank |
|---|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 77.4 | #99 | |
| 2 | Gemini 2.5 Pro | 72.2 | #276 | |
| 3 | Qwen 3.5 Omni Plus | 66.6 | #254 | |
| 4 | Qwen3.5 Omni Flash | 56.8 | #602 | |
| 5 | Audio Flamingo 3 | 52.4 | ||
| 6 | Qwen 3 Omni 30B A3B (Thinking) | 49.6 | ||
| 7 | Qwen3 Omni 30B A3B Instruct | 46.4 | #930 | |
| 8 | MOSS-Audio-8B | Thinking | 46.4 | #2150 |
| 9 | Qwen2.5-Omni-7B | 37 | #1884 | |
| 10 | MiMo-Audio-7B | 32 | ||
| 11 | MOSS-Audio-8B-Instruct | 30.6 | #2017 | |
| 12 | Phi-4 Multimodal Instruct | 26.8 | #4154 |
No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.
Interactive version: theaggregate.ai/benchmark?slug=voicegiraffe-cascaded-captions-multi-hop-reasoning · How It Works · Data refreshed daily, snapshot 2026-11-21.