Reproduce “TAEC: Trajectory-Aware Evidence Coordination for Multi-Step Visual RAG” 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
Yalun Wu, Bingzhou Wang, Boyang Wang, Peiying Wang, Shaojie He, Yunhan Wang, Shaozu Yuan, Jiawei Wang
TAEC coordinates evidence in multi-step visual document question answering by tracking unresolved answer requirements and using them to guide page admission, memory presentation, and retained-image detail. It is training-free and leaves search planning, visual interpretation, and stopping to the acting model. Across 12 model-benchmark settings, TAEC averages 66.8% accuracy versus 62.3% for a DAG agent, on tested English document-QA setups.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.