Research map / Large language models / Agents, code and math
WebThinker: Empowering Large Reasoning Models with Deep Research Capability
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 WebThinker, a deep research agent in which large reasoning models autonomously search and navigate the web via a Deep Web Explorer and draft reports during reasoning, further trained with iterative online DPO.
Problem
Large reasoning models rely on static internal knowledge, which limits performance on knowledge-intensive tasks and on producing comprehensive research reports that require synthesizing web information; existing open-source deep search agents use predefined RAG workflows.
Approach
Lets the LRM itself invoke a Deep Web Explorer (search plus link/button clicking, driven by the LRM) and writing tools (draft, check, edit) interleaved with its reasoning, and trains tool use with iterative on-policy online DPO using preference pairs ranked by correctness, tool-call count, and length.
Key results
- WebThinker-32B-RL reaches 48.5 on GAIA and 46.5 on WebWalkerQA, versus 39.8 and 34.1 for Search-o1-32B (32B models).
- WebThinker-32B-RL scores 15.8 on HLE, above Search-o1-32B (10.8) and o3-mini (High) per the paper's comparison.
- On Glaive report generation, WebThinker scores 8.0 overall (8.1 for RL version in ablation), above Gemini Deep Research at 7.9.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use WebThinker-32B-RL instead of Search-o1-32B | QwQ-32B · GAIA | Pass@1 (Avg.): 39.8 -> 48.5 | Helped |
| Use WebThinker-32B-RL instead of Search-o1-32B | QwQ-32B · WebWalkerQA | Pass@1 (Avg.): 34.1 -> 46.5 | Helped |
| Use WebThinker-32B-RL instead of Search-o1-32B | QwQ-32B · HLE | Pass@1 (Avg.): 10.8 -> 15.8 | Helped |
| Add iterative online DPO training (ablation) | WebThinker-32B-Base · Complex Problem-Solving | Avg.: 42.1 -> 45.4 | Helped |
| Use iterative online DPO instead of offline DPO (ablation) | WebThinker · Complex Problem-Solving | Avg.: 43.2 -> 45.4 | Helped |
| Add link clicking (ablation) | WebThinker-32B-RL · Complex Problem-Solving | Avg.: 42.6 -> 45.4 | Helped |
| Add Deep Web Explorer (ablation) | WebThinker-32B-RL · Complex Problem-Solving | Avg.: 38.3 -> 45.4 | Helped |
| Add iterative online DPO training (ablation) | WebThinker-32B-Base · Glaive | Avg.: 8.0 -> 8.1 | No clear effect |
| Add Deep Web Explorer (ablation) | WebThinker-32B-RL · Glaive | Avg.: 7.7 -> 8.1 | Helped |
| Add report check and edit tools (ablation) | WebThinker-32B-RL · Glaive | Coherence: 6.9 -> 7.9 | Helped |
| Add autonomous report drafting (ablation) | WebThinker-32B-RL · Glaive | Avg.: 6.6 -> 8.1 | Helped |
Surprising findings
- Standard RAG often underperforms direct reasoning or shows inconsistent gains on HLE, and removing autonomous drafting causes the largest report-quality drop.
Limitations
- Cannot process multimodal information such as images and videos.
- Supports only a limited set of tools.
- Text-only web exploration without GUI-based interaction.
Experimental setup
Datasets: SuperGPQA WebWalkerQA OpenThoughts NaturalReasoning NuminaMath glaiveai/reasoning-v1-20m GPQA GAIA HLE
Benchmarks: GPQA GAIA WebWalkerQA Humanity's Last Exam Glaive
Metrics: Pass@1 accuracy average report score (completeness, thoroughness, factuality
Baselines: Search-o1-32B RAG-QwQ-32B RAG-Qwen2.5-32B QwQ-32B DeepSeek-R1-671B OpenAI Deep Research Gemini2.0 Deep Research Grok3 DeeperSearch
Builds on: Search-o1 DPO chain-of-thought
Closest papers by meaning
- WebWatcher: Breaking New Frontiers of Vision-Language Deep Research Agent (ICLR 2026 · similarity 0.88)
- WebWeaver: Structuring Web-Scale Evidence with Dynamic Outlines for Open-Ended Deep Research (ICLR 2026 · similarity 0.88)
- Empowering Efficiency and Efficacy in WebAgent via Enabling Info-Rich Seeking (ICLR 2026 · similarity 0.87)
- SmartThinker: Progressive Chain-of-Thought Length Calibration for Efficient Large Language Model Reasoning (ICML 2026 · similarity 0.87)
- HierSearch: A Hierarchical Enterprise Deep Search Framework Integrating Local and Web Searches (AAAI 2026 · similarity 0.87)
- Trade-offs in Large Reasoning Models: An Empirical Analysis of Deliberative and Adaptive Reasoning over Foundational Capabilities (AAAI 2026 · 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.