IHBench - Task Fulfillment: 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 judge prefers the model's turn over GPT-4o Audio's for advancing the workflow, so 50 means parity with that reference; higher is better.
Metric: Win rate against GPT-4o Audio (%; share of 428 interruption points where a GPT-5.4-mini judge (high reasoning) prefers the model's next turn for advancing the workflow, 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) | 58.6 | #452 (Gemini 2.5 Flash) |
| 2 | Gemini 3.1 Pro (Preview) | 58.2 | #91 |
| 3 | Gemini 2.5 Pro | 52.6 | #262 |
| 4 | Gemma 4 12B (Reasoning) | 51.1 | #927 (Gemma 4 12B) |
| 5 | Gemma 4 12B (Non-reasoning) | 50.5 | #927 (Gemma 4 12B) |
| 6 | Gemini 2.5 Flash (Non-reasoning) | 48.8 | #452 (Gemini 2.5 Flash) |
| 7 | Voxtral-Small-24B-2507 | 30.8 | #1291 |
| 8 | Qwen3 Omni 30B A3B Instruct | 30.4 | #902 |
| 9 | Qwen2.5-Omni-7B | 18.1 | #1797 |
| 10 | Phi-4 Multimodal Instruct | 10.4 | #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-task-fulfillment · How It Works · Data refreshed daily, snapshot 2026-11-12.