Generation as Auxiliary Supervision: Enhancing Visual Understanding at Zero Inference Overhead via Decoupled Embedding Prediction

Benchmark Model Rank Results
chart-question-answering-on-charxiv-dqGAS (4B)–Accuracy: 78.4
chart-question-answering-on-charxiv-rqGAS (4B)–Accuracy: 40.5
logic-puzzle-solving-on-zebralogicGAS (2B)–Accuracy: 20.8
multimodal-reasoning-on-mathvistaGAS (2B after pretraining)–Accuracy: 71.2
multimodal-reasoning-on-mmmuGAS (4B)–Accuracy: 57.0
multimodal-reasoning-on-visulogicGAS (2B)–Accuracy: 28.6
video-question-answering-on-mvbenchGAS (2B after pretraining)–Avg.: 54.6
visual-question-answering-vqa-on-blinkGAS (4B)–Accuracy: 56.9
visual-question-answering-vqa-on-countbenchqaGAS (4B)–Accuracy: 90.8
visual-question-answering-vqa-on-realworldqaGAS (4B)–Accuracy: 70.3
zero-shot-video-question-answer-on-video-mme-w-o-subsGAS (4B)–Accuracy (%): 54.4