IHBench - Recovery Quality: leaderboard
IHBench (Salimi et al., 2026) interrupts spoken-dialogue models mid-utterance at 428 scripted points in 45 structured-workflow voice conversations across ten domains, with the user's turn given as audio; a GPT-5.4-mini judge (high reasoning) scores the model's next turn, averaged over three epochs. Score is the percentage of interruptions where the turn meets every type-specific recovery criterion (handling the cut-in without restarting, repeating or losing its place); higher is better.
Metric: Recovery pass rate (%; share of 428 interruption points meeting every type-specific recovery criterion, GPT-5.4-mini judge (high reasoning), audio input, mean of three epochs). Source: arxiv.org. Saturation forecast: Around February 2027. 27 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Flash (Thinking) | 70.4 | #452 (Gemini 2.5 Flash) |
| 2 | Gemini 2.5 Pro | 69.5 | #262 |
| 3 | Gemini 2.5 Flash (Non-reasoning) | 67.9 | #452 (Gemini 2.5 Flash) |
| 4 | Qwen3 Omni 30B A3B Instruct | 67.6 | #902 |
| 5 | Gemini 3.1 Pro (Preview) | 64.9 | #91 |
| 6 | Voxtral-Small-24B-2507 | 59.3 | #1291 |
| 7 | Gemma 4 12B (Reasoning) | 55 | #927 (Gemma 4 12B) |
| 8 | Gemma 4 12B (Non-reasoning) | 54 | #927 (Gemma 4 12B) |
| 9 | Qwen2.5-Omni-7B | 53 | #1797 |
| 10 | Phi-4 Multimodal Instruct | 46.5 | #3914 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=ihbench-recovery-quality · How It Works · Data refreshed daily, snapshot 2026-11-12.