LG AI Research 588

Recent evolution of LLMs is rapidly shifting from simple instruction-following models to Large Reasoning Models (LRMs) that solve complex problems. LRMs go through human-like reasoning processes, demonstrating PhD-level problem-solving abilities across various fields of science and technology such as physics, chemistry, and biology.

However, when information gaps occur during the reasoning process, LRMs may suffer from Hallucination, relying solely on internal knowledge, or Over-thinking, unnecessarily prolonging reasoning.

To address these issues, Agentic LRM, which actively utilizes external knowledge, is emerging as a new alternative. In this approach, when the model detects a knowledge gap during reasoning, it supplements its own knowledge through external sources. Representative examples include OpenAI's ChatGPT (Deep Research), Google DeepMind's Gemini, and LG AI Research's ChatEXAONE.

This article introduces WebDancer from Alibaba's Tongyi Lab, which strengthens LRMs' long-horizon reasoning through web search tools. WebDancer proposes a synthetic data generation methodology designed to enable models to naturally learn exploration-based reasoning and knowledge supplementation strategies.

The core methodology is as follows:

• CRAWLQA: A multi-hop dataset generation method for training the ability to explore multiple web pages to gather and evaluate information
• E2HQA: A synthetic QA dataset generation method designed for complex reasoning and external knowledge supplementation