VisualTextTrap (LLaVA-Video) - Hallucination Resistance: leaderboard

Text overlay-induced hallucination in video question answering on VisualTextTrap (2026), built on LLaVA-Video videos with on-screen text that agrees with, contradicts or is absent from the visual evidence. Accuracy on the text-contradictory samples, in percent; higher is better.

Metric: Hallucination resistance rate (%): accuracy on the text-contradictory samples of the VisualTextTrap videos built on LLaVA-Video (text overlays rendered on every frame under text-free, text-congruent and text-contradictory conditions, the contradictory overlays written by Claude-Sonnet-4.6), multiple-choice video QA under the default configuration of each model; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 5 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Gemini 3.1 Pro (Preview)72.8#99
2Qwen 3 VL 235B A22B50.2#485
3InternVL3.5-241B-A28B49.7#733
4Qwen 3 VL 8B Instruct27.8#1029
5Qwen 3 VL 30B A3B27.6#784

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=visualtexttrap-llava-video-hallucination-resistance · How It Works · Data refreshed daily, snapshot 2026-11-21.