Research map / Large language models / Retrieval and knowledge
Hybrid Deep Searcher: Scalable Parallel and Sequential Search Reasoning
Dayoon Ko, Jihyuk Kim, Haeju Park, Sohyeon Kim, Dahyun Lee, Yongrae Jo, Gunhee Kim, Moontae Lee, Kyungjae Lee
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| train on HDS-QA hybrid parallel+sequential trajectories vs single-query trajectories (ablation) | HybridDeepSearcher · MuSiQue | Acc: 21.6 -> 35.1 | Improved |
| train on hybrid trajectories vs single-query trajectories (ablation) | HybridDeepSearcher · FanOutQA | Acc: 8.7 -> 20.0 | Improved |
| train on hybrid trajectories vs single-query trajectories (ablation) | HybridDeepSearcher · BrowseComp subset | Acc: 6.4 -> 13.3 | Improved |
| apply GRPO on top of the supervised model (ablation) | HybridDeepSearcher · FRAMES | Acc: 54.0 -> 57.2 | Improved |
| apply GRPO on top of the supervised model (ablation) | HybridDeepSearcher · five benchmarks | AUC (efficiency): 0.44 -> 0.42 (FRAMES) | Hurt |
| apply GRPO on top of the supervised model (ablation) | HybridDeepSearcher · BrowseComp subset | # search turns: 6.0 -> 6.7 | Hurt |
| supervised fine-tuning on HDS-QA instead of RL-trained baselines (Qwen2.5-7B-Instruct backbone) | Qwen2.5-7B-Instruct · FanOutQA | F1: 28.2 -> 37.4 | Improved |
| add HDS-QA hybrid fine-tuning to Qwen3-8B vs prompt-based Search-o1 | Qwen3-8B · FanOutQA | F1: 26.7 -> 44.1 | Improved |
| hybrid search vs best parallel baseline RAG-R1 | HybridDeepSearcher · FanOutQA | evidence coverage rate: 53.2 -> 61.0 | Improved |
| increase search turn limit from 1 to 8 | HybridDeepSearcher · BrowseComp subset | F1: 4.0 -> 15.1 | Improved |
| increase search call limit from 2 to 16 | HybridDeepSearcher · BrowseComp subset | F1: 3.4 -> 15.1 | Improved |
| single-query fine-tuned variant vs prompt-based Search-o1 (ablation) | HybridDeepSearcher single-query · FRAMES | Acc: 48.6 -> 33.8 | Hurt |
Closest papers by meaning
- AceSearcher: Bootstrapping Reasoning and Search for LLMs via Reinforced Self-Play (NeurIPS 2025 · similarity 0.89)
- Retriever-Free Retrieval-Augmented Reasoning via Corpus-Traversing MCTS (NeurIPS 2026 · similarity 0.88)
- HierSearch: A Hierarchical Enterprise Deep Search Framework Integrating Local and Web Searches (AAAI 2026 · similarity 0.88)
- BRIGHT: A Realistic and Challenging Benchmark for Reasoning-Intensive Retrieval (ICLR 2025 · similarity 0.88)
- Fathom-DeepResearch: Unlocking Long Horizon Information Retrieval and Synthesis for SLMs (ICLR 2026 · similarity 0.88)
- Accelerating Large Language Model Reasoning via Speculative Search (ICML 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
