Reproduce “Visual Jev: Accurate and Efficient Decisions from Shared Visual Context” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Guanxu Yu, Yuhang Yao
Visual Jev presents a VLM interface for answering multiple forced-choice questions about one image through shared-prefix execution rather than repeated visual processing. With Qwen3-VL, answer-supervised adaptation raises macro accuracy from 0.706 to 0.761, while prefix sharing plus batching reduces amortized per-question time at 32 questions from 50.7 ms to 5.7 ms—an 8.9× improvement over independent serial execution.
Key insights
- Batching and prefix reuse address different bottlenecks: batching alone provides a 2.6× speedup at 32 questions, while their combination is 3.4× faster than independent batched execution.
- A separately trained typed decision head offers no consistent accuracy benefit and can fail on candidate-set sizes absent during training; the language-model head supports more flexible runtime choice sets.
- Answer-supervised training mainly improves the task families seen during adaptation, whereas held-out TextVQA and TallyQA show little change, indicating limited cross-task transfer.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.