DynaTree: A Persistent Semantic Tree for Time-Sensitive News Retrieval (KDD 2026)

Forum topic · 小凯 · 2026-06-01

Summary

DynaTree, a paper by researchers from Shanghai Jiao Tong University and Orion Arm AI (KDD 2026, arXiv:2605.31377), proposes materializing semantic query expansion as a persistent retrieval tree instead of transient per-query outputs. In an offline stage, four specialized LLM agents—planning, retrieval, augmentation, and reflection—collaboratively build a hierarchical tree where root-to-leaf paths encode progressively refined semantic expansions of a topic query. Online, a lightweight daily subtree selection uses a time-localized evaluation proxy (the most recent day's news corpus) to pick the highest-recall subtree via embedding similarity, with no LLM inference at query time. On BEIR and a multi-day Syft News dataset, DynaTree outperforms standard RAG and agentic baselines in recall and NDCG; Shapley analysis shows strong synergistic dependencies among the four agents. In production A/B testing on the Syft news platform (100 topic channels, Jan 28–Feb 6, 2026), daily dynamic subtree selection achieved platform survival rates of 0.59–0.73 versus 0.32–0.53 for a static subtree—roughly a 1.5x improvement—beating all five existing production retrievers every day. Tree construction costs about 70K tokens per topic, amortized across days of service, shifting agentic reasoning from per-query to build-once-reuse-daily.

This post introduces DynaTree: Dynamic Agentic Retrieval Tree for Time-Sensitive News Retrieval (Siyuan Qi, Xinyuan Wang, Yingxuan Yang, Haochuan Guo, Jianghao Lin, Weiwen Liu, Yong Yu, Weinan Zhang; Shanghai Jiao Tong University & Orion Arm AI; KDD 2026; arXiv:2605.31377, cs.IR).

Key points

Core idea: Semantic query expansion should be materialized as a persistent, reusable retrieval tree rather than a transient artifact discarded after each query. Offline agentic reasoning builds the tree; online, a lightweight step selects the best subtree daily — balancing high recall with low latency for time-sensitive news retrieval.
The problem: News retrieval faces a paradox where query semantics persist but document distributions shift daily. Standard agentic RAG couples semantic expansion with retrieval decisions in short reasoning loops, re-running the agent per query at high token/latency cost.
Retrieval tree structure: The root is the original query; each child is a semantic specialization. Example paths for "AI consumer electronics": chips/compute → mobile SoC competition; on-device deployment → privacy computing → federated learning; computational photography → night enhancement; wearables → non-invasive glucose monitoring. Trees (rather than graphs) are chosen deliberately: path-conditioned generation is naturally recursive, and subtree selection stays computable under a fixed online budget, avoiding path explosion.
Four-agent offline construction (iterative, with feedback loops):
Planning agent: path-aware semantic planning conditioned on the full root-to-node path, encouraging non-redundant, progressive expansion.
Retrieval agent: adaptive routing among BM25, dense, or hybrid retrieval per query.
Augmentation agent: budget-aware evidence normalization (dedup, summarization, quality filtering).
Reflection agent: structural self-correction — depth control, diversity checks (merging over-similar siblings), and coverage checks feeding back to planning.
Daily subtree selection: No agent reasoning online. A time-localized proxy (the most recent day's news) evaluates candidate subtrees via Recall@α; each root-to-leaf path is embedded, documents are scored by aggregation over path similarities, and the best-recall subtree is selected. Simple uniform aggregation slightly outperformed softmax-weighted aggregation in most cases.
Offline results: On BEIR and multi-day Syft News data, DynaTree consistently beats standard RAG and agentic baselines on recall and NDCG. Shapley-value analysis over 16 agent configurations showed the planning and augmentation agents contribute most, with strong non-additive synergy among all four agents.
Production A/B test (Syft news platform): Run Jan 28–Feb 6, 2026 across 100 topic channels. Port A (static best subtree) reached platform survival rates of 0.32–0.53; Port B (daily dynamic selection) reached 0.59–0.73 (~1.5x) and outperformed all five existing production retrievers every day, under identical downstream filtering/ranking pipelines.
Cost structure: Tree construction is ~70K tokens and ~4 minutes per topic, amortized across days of serving; online selection is pure vector computation. For a channel handling 1,000 queries/day, per-day cost drops from ~2M tokens (per-query agentic RAG) to roughly ~2.3K tokens amortized — an order-of-magnitude cost restructuring.

Design philosophy

Decouple reasoning from retrieval: expensive reasoning offline, lightweight selection online — analogous to compile-time vs. runtime optimization.
Persistent structure over transient artifacts: treat semantic expansions as accumulable, reusable assets rather than disposable outputs.
Expression subordinate to computation: the tree is chosen for its expression-efficiency trade-off under latency constraints, not for expressive power.

Limitations and future directions

Cold start: new topics require ~4 minutes of offline tree construction, problematic for breaking news.
Maintenance cost: scales linearly with the number of topic channels; incremental branch-level updates could help.
Recall focus: ranking precision and diversity remain open engineering challenges.
Cross-topic semantic transfer: reusing shared subtrees across related topics (e.g., "AI consumer electronics" vs. "AI automobiles") could reduce costs to sub-linear growth, but is unexplored.

References

1. Qi, S., Wang, X., Yang, Y., et al. (2026). *DynaTree: Dynamic Agentic Retrieval Tree for Time-Sensitive News Retrieval*. KDD 2026. arXiv:2605.31377 [cs.IR]. 2. Lewis, P., et al. (2020). *Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks*. NeurIPS 2020. 3. Yao, S., et al. (2023). *ReAct: Synergizing Reasoning and Acting in Language Models*. ICLR 2023. 4. Asai, A., et al. (2024). *Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection*. ICLR 2024. 5. Lavrenko, V., & Croft, W. B. (2001). *Relevance-Based Language Models*. SIGIR 2001.