TL;DR
DynaTree is a two-stage retrieval framework that decouples expensive agentic reasoning from real-time execution to handle time-sensitive news retrieval. By building a persistent, structured "retrieval tree" offline and using a lightweight proxy to dynamically select the best subtrees online, it achieves state-of-the-art recall and temporal adaptability at a fraction of the operational cost of standard agentic RAG.
What Problem Does This Paper Address?
Retrieval-Augmented Generation (RAG) systems increasingly rely on "agentic" loops—where large language models (LLMs) iteratively plan, rewrite queries, and evaluate results—to improve retrieval quality. However, these systems usually couple semantic expansion directly with per-query online inference. This creates a massive bottleneck for time-sensitive applications like news retrieval, where latency must be low, token costs must be controlled, and recall must be exceptionally high.
Furthermore, news retrieval suffers from intense temporal drift. While a user's high-level interest (e.g., "AI in consumer electronics") might remain stable, the specific subtopics, breaking stories, and relevant articles shift daily. Traditional agentic RAG discards its reasoning after each query, leading to redundant computation and an inability to model how evolving subtopics relate to one another over time.
Methods
DynaTree solves this by splitting the retrieval process into two distinct stages:
- Stage I: Agentic Retrieval Tree Construction (Offline): Instead of executing an agent loop for every single user query, DynaTree amortizes this cost by building a reusable, hierarchical retrieval tree for each broad query topic. Specialized agents work together recursively:
- Planning Agent: Proposes queries based on the path context to explore deeper subtopics.
- Retrieval Agent: Dynamically routes queries to the best tool (BM25, dense, or hybrid retrieval).
- Augmented Agent: Normalizes and deduplicates retrieved evidence under a strict token budget.
- Reflection Agent: Evaluates the new branches and applies structural corrections (Add, Prune, Revise) to prevent semantic drift.
- Stage II: Dynamic Adaptive Subtree Selection (Online/Daily): In production, the tree structure remains fixed, but the system adapts to daily news shifts by selecting the most relevant subtree. It constructs a compact "evaluation proxy" of documents, uses an LLM to weakly label them for relevance, and optimizes for the subtree that maximizes recall.
At query time, documents are scored using a path-ensemble formulation that aggregates similarity across all paths in the selected subtree, favoring articles that receive consistent support from multiple semantic angles.
Datasets
The authors evaluated DynaTree using two primary sources:
- Syft News Dataset: A newly introduced, large-scale proprietary benchmark consisting of daily snapshots of real-world news streams (across seven distinct dates in June 2025). It includes 50 core query themes, alongside complex and long-tail themes designed to simulate realistic temporal distribution shifts.
- BEIR Benchmark: Five public datasets (Trec-Covid, NfCorpus, FiQA, SciFact, and SciDocs) were used to evaluate how well the framework generalizes to static, out-of-domain retrieval tasks.
Key Insights and Ideas
- Semantic Expansion as a Persistent Asset: Rather than treating query rewrites as transient, throwaway artifacts, structuring them into a persistent tree allows the system to accumulate and reuse semantic knowledge.
- Decoupling Exploration from Optimization: You don't need to re-run expensive LLM planning loops to adapt to today's news. By separating the discovery of semantic paths (Stage I) from the selection of active paths (Stage II), the system achieves the quality of agentic RAG with the speed of standard dense retrieval.
- Relative Ordering via Compact Proxies: Stage II doesn't need to calculate absolute recall across the entire corpus. Evaluating candidate subtrees against a small, diversity-aware proxy dataset (around 150–200 samples) is sufficient because subtree selection only relies on preserving the relative ranking of the subtrees.
Key Results
- Offline Performance: On the Syft News benchmark, DynaTree consistently achieved the highest Recall@100 and NDCG@10 across all daily snapshots, outperforming strong baselines like ReAct, Reflexion, GraphRAG, and HyDE. It also achieved the best average performance across the static BEIR datasets.
- Online Production Deployment: DynaTree was directly deployed in the production pipeline of the Syft news platform. During live A/B testing (Jan. 28 to Feb. 6, 2026), the dynamically adapted variant (Port B) boosted the platform's survival rate (the fraction of retrieved news passing downstream quality control) from 0.32–0.53 up to 0.59–0.73. This represented a ~1.5× gain over using a fixed, static subtree and outperformed all existing production recallers every single day.
Limitations and Future Work
While DynaTree excels at handling predictable topic channels, it relies on the tree structure generated in Stage I. For entirely unprecedented, "zero-day" breaking news topics that fall completely outside the pre-built tree's semantic scope, the system must fall back to traditional low-latency recallers. The authors note that in these scenarios, tree updates are queued asynchronously via the Reflection Agent to avoid blocking live user traffic, but true real-time structural expansion for entirely novel concepts remains an area for future optimization.
Summary updated