DynaTree: A Two-Stage Agentic Framework for Time-Sensitive News Retrieval
A recent research publication presents DynaTree, a two-phase framework designed for efficient and adaptive retrieval of news. This system tackles the shortcomings of current agentic Retrieval-Augmented Generation (RAG) techniques, which frequently link semantic expansion with retrieval choices in short-term inference cycles, resulting in elevated costs and inadequate performance for time-sensitive news. DynaTree functions in two phases: the first is an offline phase where coordinated agents build a reusable retrieval tree that embodies the semantic landscape of a query topic, followed by an online phase that executes lightweight daily subtree selection based on a time-localized evaluation proxy, without additional agentic reasoning, tree alterations, or retraining. Tests on a multi-day Syft news benchmark and various BEIR datasets reveal that DynaTree consistently surpasses standard RAG and previous agentic approaches in recall and ranking performance.
Key facts
- DynaTree is a two-stage framework for time-sensitive news retrieval.
- It uses coordinated agents in an offline stage to build a reusable retrieval tree.
- The online stage performs lightweight daily subtree selection without further reasoning.
- Experiments were conducted on a multi-day Syft news benchmark and multiple BEIR datasets.
- DynaTree outperforms standard RAG and prior agentic methods in recall and ranking.
- The paper is available on arXiv with ID 2605.31377.
- Existing agentic RAG methods couple semantic expansion with retrieval decisions in short-horizon loops.
- DynaTree reduces inference cost by avoiding agentic reasoning during online retrieval.
Entities
Institutions
- arXiv