Reproduce “ALDEN: Reinforcement Learning for Active Navigation and Evidence Gathering in Long Documents” 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

Tianyu Yang, Terry Ruas, Yijun Tian, Jan Philip Wahle, Daniel Kurzawe, Bela Gipp

ALDEN trains a vision-language model to navigate long documents by choosing to search, fetch a page by index, or answer. Across five benchmarks, its ColQwen plus ColBERT configuration has average GPT-4o-judged accuracy of 0.446, versus 0.415 for MDocAgent with the same retrievers, a 7.47% relative improvement. A targeted DUDE-sub test supports fetch in that setting, not universal benefit or efficiency.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.