Research map / Multi-agent systems and games / LLM agents

HierSearch: A Hierarchical Enterprise Deep Search Framework Integrating Local and Web Searches

AAAI 2026 · Multi-agent systems and games › LLM agents · approach: Agent data synthesis and behavior studies · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes HierSearch, a hierarchical RL-trained deep search framework where local and Web deep search agents feed evidence through a reasoning-aware knowledge refiner to a planner agent, outperforming flat RL and multi-source RAG baselines on six general, financial, and medical QA benchmarks.

Problem

Enterprise deep search must use both private local knowledge sources and the Web, but flat RL over many search tools has low training efficiency and under-explores harder Web tools.

Approach

Train separate low-level local and Web deep search agents with GRPO, then train a high-level planner that calls them as tools; a reasoning-aware knowledge refiner filters evidence by embedding similarity to subsequent thinking and to the agents' conclusions before passing it to the planner.

Key results

• HierSearch achieves 53.00 EM / 62.83 F1 on MuSiQue and 49.94 EM on BioASQ, the best across all six benchmarks versus baselines.
• Flat-RL variant HierSearch w/o HRL reaches only 23.50% web search success on NQ versus 81.25% for HierSearch.
• Ablation removing the local agent drops MuSiQue EM from 53.00 to 29.75 (23.25% relative drop).

Surprising findings

• Flat RL reached only 18.5% of tool calls on Web search at step 290, showing under-exploration of the harder knowledge source.
• Search-o1 and WebThinker consume far more reasoning tokens (1,500 to 4,277) without matching accuracy gains.

Limitations

• Evaluation uses randomly sampled subsets (373/340/400 samples) of test sets rather than full benchmarks.
• Local knowledge bases for general and medical domains are sampled corpora, not full Wikipedia/PubMed.

Experimental setup

Datasets: MuSiQue OmniEval BioASQ NQ HotpotQA PubmedQA Wikipedia dump PubMed dump

Benchmarks: MuSiQue OmniEval BioASQ NQ HotpotQA PubmedQA

Metrics: EM F1 average local and Web search calls search success rate reasoning success rate latency

Baselines: DeepSeek-R1 HippoRAG R1-Searcher ReCall DeepResearcher Search-o1 WebThinker HM-RAG PrefRAG CRAG HierSearch w/o HRL

Builds on: GRPO HippoRAG ReCall Search-o1 WebThinker

Closest papers by meaning

• Fathom-DeepResearch: Unlocking Long Horizon Information Retrieval and Synthesis for SLMs (ICLR 2026 · similarity 0.90)
• Search More, Think Less: Rethinking Long-Horizon Agentic Search for Efficiency and Generalization (NeurIPS 2026 · similarity 0.89)
• Empowering Efficiency and Efficacy in WebAgent via Enabling Info-Rich Seeking (ICLR 2026 · similarity 0.89)
• Quest: Training Frontier Deep Research Agents with Fully Synthetic Tasks (NeurIPS 2026 · similarity 0.88)
• MM-DeepResearch: A Simple and Effective Multimodal Agentic Search Baseline (ICML 2026 · similarity 0.88)
• OpenSearch-VL: An Open Recipe for Frontier Multimodal Search Agents (NeurIPS 2026 · similarity 0.88)