Reproduce “ReAccess: Look Once, Stay Focused with Efficient Test Time Memory” 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
Tianhao Chen, Yilong Fan, Yuheng Wu, Xiaobin Hu, Dongman Lee
ReAccess uses question-conditioned associative memory to preserve visual evidence during image prefill and retrieve it as a VLM generates an answer, without re-encoding the image. On V* Spatial, Qwen2.5-VL-7B reaches 82.9% accuracy, versus 77.6% for the frozen backbone and 76.3% for LVR. Evidence masking lowers accuracy to 80.3%, supporting a role for memory evidence, but ReAccess does not lead every task.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.