Research map / Large language models / Agents, code and math
Enhancing Agentic Search via Data Synthesis on Hierarchical Constraint Satisfaction
Ziyi Xia, Kun Luo, Hongjin Qian, Siqi Bao, Zheng Liu
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 InfoSeek data (SFT then GRPO) instead of search-agent baselines trained on NQ/HotpotQA | Qwen2.5-3B-Instruct · 7 QA benchmarks (average) | exact match: 40.6 -> 42.7 (vs InForage-3B) | Improved |
| InfoSeeker-3B vs baselines on single-hop NQ and TQA | InfoSeeker-3B · NQ, TriviaQA | exact match: NQ 41.7 vs AutoRefine 43.6; TQA 56.1 vs ZeroSearch 61.5 | Mixed |
| InfoSeeker-3B vs strongest baselines on multi-hop sets | InfoSeeker-3B · HotpotQA, 2Wiki | exact match: HQA 40.9 -> 44.6; 2Wiki 42.8 -> 50.0 | Improved |
| train 3B model on InfoSeek | InfoSeeker-3B · BrowseComp-Plus | accuracy: 3.9 -> 15.3 (vs SearchR1-32B) | Improved |
| InfoSeeker-3B vs GPT-5 on BrowseComp-Plus | InfoSeeker-3B · BrowseComp-Plus | accuracy: 55.9 vs 15.3 | Hurt |
| add InfoSeek SFT stage before RL (ablation) | Qwen2.5-3B-Instruct · 7 QA benchmarks | exact match | Improved |
| scale model from 3B to 7B (ablation) | InfoSeeker · 7 QA benchmarks | exact match | Improved |
| train on InfoSeek instead of only NQ+HotpotQA (ablation) | InfoSeeker-3B · BrowseComp-Plus | accuracy and search calls | Improved |
| restrict InfoSeek to fewer than five vertices (ablation) | InfoSeeker-3B · BrowseComp-Plus | accuracy and search calls | Improved |
| increase training data scale (1K, 5K, 18K) (ablation) | InfoSeeker-3B · BrowseComp-Plus | accuracy and search calls | Improved |
| increase vertices per question (ablation) | InfoSeek dataset · InfoSeek difficulty | failure rate of Qwen2.5-72B: 88.1 -> 94.1 | Improved |
Closest papers by meaning
- Empowering Efficiency and Efficacy in WebAgent via Enabling Info-Rich Seeking (ICLR 2026 · similarity 0.90)
- InfoMosaic-Bench: Evaluating Multi-Source Information Seeking in Tool-Augmented Agents (ICLR 2026 · similarity 0.90)
- Mind2Web 2: Evaluating Agentic Search with Agent-as-a-Judge (NeurIPS 2025 · similarity 0.89)
- Fathom-DeepResearch: Unlocking Long Horizon Information Retrieval and Synthesis for SLMs (ICLR 2026 · similarity 0.89)
- MindSearch: Mimicking Human Minds Elicits Deep AI Searcher (ICLR 2025 · similarity 0.89)
- WideSearch: Benchmarking Agentic Broad Info-Seeking (ICLR 2026 · similarity 0.88)
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.
