MSI-Bench - Mandarin Selective Disclosure: leaderboard

Metric: All-pass rate (%; the 96 selective-disclosure cases, where earlier speech says which facts to withhold from which listeners and those listeners later ask for them, 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 2029. 15 models tracked.

Top models

#ModelScore
1Gemini 3.1 Pro (Preview)53.1
2GPT Audio39.6
3Phi-4 Multimodal Instruct36.5
4Gemini 3.5 Flash35.4
5Gemma 4 12B (Non-reasoning)27.1
6GPT Audio Mini20.8
7Qwen2.5-Omni-7B16.7
8Gemma 4 12B (Reasoning)13.5

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