PointThinker: Point-Incentivized Parallel Thinking for Multimodal Large Language Model

Benchmark Model Rank Results
mathematical-reasoning-on-mathversePointThinker-7B (extra data)–Accuracy: 45.9
multimodal-reasoning-on-mathvisionPointThinker-7B (extra data)–Accuracy: 28.1
multimodal-reasoning-on-mathvistaPointThinker-7B (extra data)–Accuracy: 74.2
multimodal-reasoning-on-mmmuPointThinker-7B (extra data)–Accuracy: 59.3
spatial-reasoning-on-vsi-benchPointThinker-7B (extra data)–Accuracy: 37.7
video-question-answering-on-mmvuPointThinker-7B (extra data)–Accuracy: 66.1
video-question-answering-on-mvbenchPointThinker-7B (extra data)–Avg.: 66.2
video-question-answering-on-tempcompass-overallPointThinker-7B (extra data)–Accuracy: 74.7
video-question-answering-on-videommmuPointThinker-7B (extra data)–Accuracy: 53.0
vision-language-understanding-on-mmstarPointThinker-7B (extra data)–Accuracy: 66.8
zero-shot-video-question-answer-on-video-mme-w-o-subsPointThinker-7B (extra data)–Accuracy (%): 62.5