WebThinker: Empowering Large Reasoning Models with Deep Research Capability
- Paper: WebThinker: Empowering Large Reasoning Models with Deep Research Capability (arXiv, April 2025)
- Authors: Xiaoxi Li, Jiajie Jin, Guanting Dong, Hongjin Qian, Yongkang Wu, Ji-Rong Wen, et al. (8 authors)
- Category: Deep Research
- Problem: Large reasoning models (LRMs) are strong at chain-of-thought reasoning but rely on static, internal knowledge, which limits their performance on complex, knowledge-intensive tasks that require up-to-date, real-world information.
- Proposed system: WebThinker enables LRMs to autonomously perform deep research — searching the web, clicking into and navigating relevant pages, extracting useful content, and drafting a final report — all interleaved with their reasoning process.
- Deep Research Strategy: For complex questions, the system decomposes the task into multiple research phases, iteratively searching, diving into web pages, and synthesizing findings before producing the answer or report.
- Training pipeline: 1. Synthesize reasoning-search interaction trajectories for supervised fine-tuning; 2. Apply online reinforcement learning with Direct Preference Optimization (DPO) to further optimize both the model's reasoning behavior and its web search/browsing actions.
- Results: According to the paper, WebThinker outperforms prior baselines on challenging benchmarks spanning complex reasoning and knowledge-seeking tasks, demonstrating the value of tightly coupling reasoning with web interaction rather than using a rigid retrieval pipeline.
- The classic search stack (recall → ranking → generation) remains dominant, but agentic approaches make retrieval policy a learnable behavior.
- Evaluation is shifting from static nDCG toward task success rate, citation accuracy, and multi-hop reasoning chain quality; LLM-as-judge should be cross-validated with human evaluation.
- Real-world deployment must account for latency, cost, hallucination risk, and safety when browsing the open web (e.g., source filtering, adversarial content detection).
- A Comprehensive Survey of Deep Research (arXiv 2506.12594)
- A Survey of LLM-based Deep Search Agents (arXiv 2508.05668)
- Agentic Reasoning: A Streamlined Framework for Enhancing LLM Reasoning (arXiv 2502.04644)
- AgentIR: Reasoning-Aware Retrieval for Deep Research Agents
Key points
Context within Deep Research
The post situates WebThinker at the intersection of information retrieval, retrieval-augmented generation (RAG), and agentic search. Rather than treating retrieval as a fixed, one-shot pipeline (retrieve → rerank → generate), WebThinker makes the number and strategy of search actions themselves part of the model's reasoning, reflecting the broader shift toward agentic deep research systems.
Engineering considerations discussed in the post
Related entries
Source: https://arxiv.org/abs/2504.21776