DynaTree: Dynamic Agentic Retrieval Tree for Time-Sensitive News Retrieval
Year: 2026
Type: conference-paper
Abstract: Agentic Retrieval-Augmented Generation (RAG) improves retrieval by integrating planning, tool use, and iterative reasoning, but most existing agentic RAG methods tightly couple semantic expansion with retrieval decisions within short-horizon inference loops, resulting in high inference costs and limited suitability for time-sensitive, high-recall tasks such as news retrieval. We propose DynaTree (Dynamic Agentic Retrieval Tree), a two-stage framework in which agentic RAG is used for offline semantic exploration, while online retrieval is performed without agentic inference. In the first stage, DynaTree adopts an agentic RAG process to construct a reusable retrieval tree for each query topic, where coordinated agents plan semantic expansions, execute diverse retrieval strategies, normalize retrieved evidence, and iteratively refine the tree using retrieval feedback, materializing a persistent representation of the query's semantic space. In the second stage, DynaTree operates purely as a structure-aware retrieval system, performing lightweight daily subtree selection over a time-localized evaluation proxy, with large language models used only for rapid relevance labeling, without further agentic reasoning, tree modification, or retraining. Experiments on our multi-day Syft news benchmark and multiple BEIR datasets show that DynaTree achieves state-of-the-art recall and NDCG, consistently outperforming standard RAG and prior agentic baselines. Moreover, DynaTree has been directly deployed in the Syft production system and evaluated through online A/B testing from Jan. 28 to Feb. 6, 2026. Across all days, the dynamically adapted variant with daily subtree selection improves survival rate from 0.32-0.53 to 0.59-0.73, yielding consistent gains of up to ∼ 1.5× over a fixed offline-selected subtree and outperforming all existing production recallers on every evaluation day. These online gains persist under live user traffic, demonstrating that DynaTree translates structure-aware, agentic offline reasoning into tangible improvements in coverage, freshness, and relevance for real-world news retrieval.
Institution: Shanghai Jiao Tong University
Language: English
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7196922845