Reproduce “AdaGATE: Adaptive Gap-Aware Token-Efficient Evidence Assembly for Multi-Hop Retrieval-Augmented Generation” 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
Yilin Guo, Yinshan Wang, Yixuan Wang
AdaGATE treats multi-hop retrieval-augmented generation as an iterative, token-limited evidence-repair problem rather than a one-shot passage-selection task. By tracking missing entity–relation facts, issuing targeted and fallback queries, and replacing redundant evidence, it achieved the best supporting-document F1 across clean, noisy, and redundant retrieval settings while using roughly 2.6–3.1 times fewer input tokens than Adaptive-k on clean data.
Key insights
- Explicit gap tracking helps distinguish genuinely complementary evidence from passages that are merely similar to already selected content.
- Redundancy-aware selection can improve both efficiency and evidence quality: under redundancy injection, AdaGATE used about two passages and 220–232 tokens while achieving its strongest evidence F1.
- Answer accuracy and evidence quality do not always align; methods such as Adaptive-k can obtain high recall or favorable automated relevance scores by passing large contexts, despite very low evidence precision.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.