MSI-Bench - Mandarin Speaker Authority: leaderboard

Metric: All-pass rate (%; the 96 speaker-authority cases, where an unauthorized participant asks for an action another speaker controls and the assistant must execute only what that speaker approved, refuse, or check first, Mandarin split; a case passes when all its atomic rubrics pass, judged by DeepSeek V4 Pro on the dialogue transcript, with a deterministic tool-call validator as one more rubric for cases that need a tool call; one prediction per model and case from the audio of a multi-party, multi-turn scene ending in an assistant-directed handoff; open-weight models served with JSON-schema constrained decoding). Source: arxiv.org. Saturation forecast: Around December 2026. 15 models tracked.

Top models

#ModelScore
1Gemini 3.5 Flash81.3
2Gemini 3.1 Pro (Preview)80.2
3GPT Audio74
4GPT Audio Mini27.1
5Qwen2.5-Omni-7B16.7
6Phi-4 Multimodal Instruct1
7Gemma 4 12B (Reasoning)0
8Gemma 4 12B (Non-reasoning)0

Interactive version: theaggregate.ai/benchmark?slug=msi-bench-mandarin-speaker-authority · How It Works · Data refreshed daily, snapshot 2026-09-26.