DuplexSpeechBench-IFEval (IAS): leaderboard
DuplexSpeechBench-IFEval (2026) evaluates implicit instruction following in full-duplex voice agents over 1,038 cases across eight assistant personas, reporting the persona-only conditioning setting where the required floor behavior must be inferred from a role. This board is the Instruction Adherence Score, a deterministic measure of real-time floor management, in percent, higher is better. The persona-blind Moshi control is not ranked.
Metric: Instruction Adherence Score (%; deterministic real-time floor-management adherence under persona-only conditioning, L2). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Fun-Audio-Chat-8B | 47.1 | |
| 2 | MiniCPM-o-4.5 | 43.2 | #1048 |
| 3 | F-Actor | 25.8 | |
| 4 | gpt-realtime-2.1 | 20.6 | |
| 5 | PersonaPlex | 6.5 |
Interactive version: theaggregate.ai/benchmark?slug=duplexspeechbench-ifeval-ias · How It Works · Data refreshed daily, snapshot 2026-11-12.