MedJev scores medical image answers in one forward pass
Original title: MedJev: A System One Vision-Language Decision Model for Fast and Effective Medical Image Understanding
Built on Qwen3.5-4B, MedJev-4B reports a 7.93-point accuracy gain and 6.04 times speedup.
MedJev maps a medical image, question and candidate answers to a decision in one multimodal forward pass, skipping autoregressive reasoning and answer tokens. It drops upper language layers and the generative pathway, keeps the vision pathway and visual tokens, and scores candidates with pretrained query-key matching plus an MLP residual. On four classification and visual question-answering benchmarks, the authors say MedJev-4B is more accurate and faster than the base model and leads 4B–9B general and medical vision-language models among 20 compared systems.
Unconfirmed: Accuracy, speedup and ranking claims are reported by the authors, not independently verified.
- Inference cost no longer scales with generated response length.
- The decision head avoids parsing an answer from free-form text.
- Comparisons cover both general and medical vision-language models in the 4B–9B range.
From the original article
1 min read at huggingface.coMedical image understanding with vision-language models (VLMs) typically relies on autoregressively generating reasoning and answer tokens. These unnecessary tokens make the inference cost scale with the response length and force the answer to be parsed from free-form text, incurring avoidable latency and decision errors.
By Fan Zhang, Hao Wu, Jinyang Wu · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.