Towards Tool-augmented LLM Agents

With the rapid advancement of Artificial General Intelligence (AGI) research, large language models (LLMs) are gradually evolving from passive language generators into proactive problem solvers and task executors. In this process, the ability of LLMs to utilize external tools is widely regarded as a key bridge connecting language intelligence with real-world interaction. However, current LLM agents still face significant challenges in complex task scenarios, including limited generalization ability and low end-to-end execution accuracy in multi-turn interactions. This talk explores the challenges faced by mainstream LLM agent models in real-world applications and focuses on systematic approaches to building tool-use capabilities in LLMs. These include key technical pathways such as data synthesis, post-training, and iterative self-evolution. Additionally, the report presents cutting-edge progress in personalized tool use, encompassing methods such as user preference modeling, query completion, and proactive tool use, aiming to significantly reduce user interaction costs and improve the practicality and intelligence of LLM agents.