DuplexSpeechBench-IFEval (PAS): leaderboard
DuplexSpeechBench-IFEval (2026) evaluates implicit instruction following in full-duplex voice agents across eight personas, reporting the persona-only conditioning setting. This board is the Persona Adherence Score, a GPT-4o-judged measure of persona-consistent content on a zero to one hundred scale, higher is better. The persona-blind Moshi control is not ranked.
Metric: Persona Adherence Score (0-100; GPT-4o-judged persona-consistent content 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 | MiniCPM-o-4.5 | 66.2 | #1048 |
| 2 | Fun-Audio-Chat-8B | 65.9 | |
| 3 | gpt-realtime-2.1 | 62.6 | |
| 4 | PersonaPlex | 22.9 | |
| 5 | F-Actor | 7.6 |
Interactive version: theaggregate.ai/benchmark?slug=duplexspeechbench-ifeval-pas · How It Works · Data refreshed daily, snapshot 2026-11-12.