MedJev: A System One Vision-Language Decision Model for Fast and Effective Medical Image Understanding
MedJev reports a 7.93 percentage-point accuracy gain and a 6.04× speedup over its Qwen3.5-4B base model.
It selects among candidate answers in a single multimodal forward pass, without generating reasoning or answer text. The model removes upper language layers and the generative output pathway while retaining the full vision pathway and all visual tokens. A decision head scores the candidates using pretrained matching and a lightweight neural network refinement. The authors report improved accuracy and speed on all four medical image classification and visual question answering benchmarks tested.
HF Daily Papers' note
It selects among candidate answers in a single multimodal forward pass, without generating reasoning or answer text. The model removes upper language layers and the generative output pathway while retaining the full vision pathway and all visual tokens. A decision head scores the candidates using pretrained matching and a lightweight neural network refinement. The authors report improved accuracy and speed on all four medical image classification and visual question answering benchmarks tested.
HF Daily Papers' note
score 5