Hallusion Bench
multimodal
+
+
+
+
About
HallusionBench is an advanced diagnostic benchmark for evaluating visual-language models' susceptibility to hallucinations and visual illusions. Featuring 346 images with 1,129 questions, this benchmark tests models' image-context reasoning abilities, focusing on detecting entangled language hallucinations and visual illusions. HallusionBench provides comprehensive evaluation of multimodal models' reliability and accuracy in visual understanding tasks.
+
+
+
+
Evaluation Stats
Total Models2
Organizations1
Verified Results0
Self-Reported2
+
+
+
+
Benchmark Details
Max Score1
Language
en
+
+
+
+
Performance Overview
Score distribution and top performers
Score Distribution
2 models
Top Score
55.2%
Average Score
54.0%
High Performers (80%+)
0Top Organizations
#1Alibaba Cloud / Qwen Team
2 models
54.0%
+
+
+
+
Leaderboard
2 models ranked by performance on Hallusion Bench
| License | Links | ||||
|---|---|---|---|---|---|
| Jan 26, 2025 | tongyi-qianwen | 55.2% | |||
| Jan 26, 2025 | Apache 2.0 | 52.9% |
