LG AI Research 587

The recent evolution of LLMs is rapidly shifting beyond simple instruction-based models toward Large Reasoning Models (LRM) that perform deep reasoning. LRMs go through a long reasoning process (long-CoT) that analyzes complex problems step by step, showing outstanding performance in science and technology fields, but they also have limitations such as Hallucination and Over-thinking that occur when information gaps arise.

To address this, Agentic LRM, in which the model itself detects a lack of knowledge and utilizes external knowledge, is drawing attention. OpenAI's ChatGPT (Deep Research) and Google's Gemini are representative examples.

WebDancer, proposed by Alibaba's Tongyi Lab, is a methodology that strengthens the reasoning capabilities of LRMs by utilizing web search tools. The core idea is to generate advanced synthetic data so that the model can learn exploration-based reasoning and information-supplementation strategies.

Two key methodologies are used for data generation:

• CRAWLQA: a Multi-hop QA dataset that requires exploring multiple sub-pages and synthesizing information
• E2HQA: an Easy-to-Hard approach that restructures questions starting from simple ones to progressively require more complex reasoning

WebDancer is designed based on the ReAct (Reasoning + Acting) structure, learning Thought–Action–Observation trajectories in which the model thinks and acts step by step and reflects the results back into its reasoning.