Beyond Correctness: Benchmarking and Aligning Response Behaviors in Hybrid-Thinking MLLMs

Benchmark Model Rank Results
chart-question-answering-on-chartqaQwen3-VL-4B + PatternRL–1:1 Accuracy: 85.42
infographic-visual-question-answering-on-infovqaQwen3-VL-8B + PatternRL–ANLS: 88.05
mathematical-reasoning-on-mathverseQwen3-VL-4B + PatternRL–Accuracy: 70.21
multimodal-reasoning-on-mathvistaQwen3-VL-8B + PatternRL–Accuracy: 86.61
multimodal-reasoning-on-visulogicQwen3-VL-8B + PatternRL–Accuracy: 25.38
visual-question-answering-on-docvqa-testQwen3-VL-8B + PatternRL–ANLS: 0.9254