Sens-VisualNews: leaderboard

Sens-VisualNews (2026) tests whether multimodal models can tell if a news image is sensational, meaning built to provoke shock, fear, anger or disgust, over 9,576 VisualNews images balanced between the two classes. Models answer yes or no zero-shot, and the score is top-1 accuracy in percent on the test split (chance 50); higher is better.

Metric: Top-1 accuracy (%) on the test split of the full dataset (majority-vote labels of three annotators), zero-shot yes or no answer to whether the news image is sensational, greedy decoding, each model with the prompt and prompt format that scored best on the development split; balanced classes, chance 50; higher is better. Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 12 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1LLaVA-OneVision-1.5-8B87.6#1864
2Qwen 3 VL 8B87#917
3LLaVA-OneVision-1.5-4B86.7#1855
4Qwen 3 VL 2B86.3#2134
5Qwen 3 VL 4B85#1151
6InternVL3.5-8B83.8#1604
7InternVL3.5-4B83.4#1930
8LLaVA-OneVision-7B81.5#2981
9LLaVA-OneVision-0.5B81.3#7956
10InternVL3.5-2B77.9#3284
11InternVL3.5-1B76.1#6112
12SmolVLM2 2.2B62.2

Interactive version: theaggregate.ai/benchmark?slug=sens-visualnews · How It Works · Data refreshed daily, snapshot 2026-11-21.