Reproduce “When Deep Research Agents Stagnate: Enhancing Reasoning with Retrieval-Aware Agent Control” 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
Heydar Soudani, Elizabeth Lingg, Faegheh Hasibi, Navid Rekabsaz
RAAC controls deep research agents by tracking document novelty, criteria coverage, query diversity, and similarity to the original question, then choosing to continue, redirect a stale but recoverable search, or stop. Across seven agents and three datasets, WebWeaver’s BrowseComp-Plus recall rose from 64.9% to 67.2% and accuracy from 47.8% to 51.3%, while calls fell from 75.0 to 50.7, with exceptions.
Key insights
- A systematic analysis revealed 'reasoning stagnation' as a consistent failure mode across diverse DRAs, where intermediate recall and accuracy quickly saturate after initial steps.
- DRAs struggle with effective exploration, increasingly revisiting previously seen documents and failing to discover new relevant information over time.
- Unsupervised, trajectory-based signals, particularly Document Novelty, effectively inform an external controller about an agent's progress and can guide its adaptive actions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.