TraceAV-Bench - Multimodal Hallucination: leaderboard

The multimodal hallucination dimension of TraceAV-Bench (2026): deceptive questions that plant a false cross-modal or temporal premise in long audio-visual videos (visual-to-audio and audio-to-visual deception, temporal splicing). The score is accuracy in percent, the mean of the three sub-tasks; higher is better.

Metric: Accuracy (%), average of the three multimodal hallucination sub-tasks (625 deceptive questions), four-option multiple choice over long audio-visual videos (10 to 140 minutes), exact match of the selected option set, each model at its official inference setting, with the full audio and video input; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 15 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Gemini 3.1 Pro (Preview)84.61#102
2Gemini 3 Flash79.81#180
3Gemini 2.0 Flash77.85#731
4Gemini 2.5 Pro76.42#280
5Qwen 3 Omni 30B A3B67.46#851
6Ming-Flash-Omni-2.067.07#1258
7Gemma 4 E4B66.6#2199
8MiniCPM-o-4.566.52#1251
9Gemini 2.5 Flash64.85#466
10Qwen2.5-Omni-7B56.5#1888
11OmniVinci-9B42.81
12video-SALMONN 240.97#2695
13HumanOmni-7B33.4#3241
14Baichuan-Omni-1.530.68#3189
15VideoLLaMA2.1-7B-AV28.26#4199

No result here: #2 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=traceav-bench-multimodal-hallucination · How It Works · Data refreshed daily, snapshot 2026-11-22.