Lumos is a cutting-edge Language Agents system with a General Agent Modular Framework that unifies a variety of interactive tasks with competitive performance. Trained with ~40K diverse high-quality subgoal/action annotations, Lumos surpasses GPT-series agents on complex QA tasks and outperforms domain-specific agents by a large margin on unseen tasks like WebShop. Lumos also exceeds larger open agents on math tasks and performs the best among baseline formulations like Chain-of-Thought Training and Integrated Agent Training. The system’s focus on high-level subgoals and planning modules sets it apart from traditional methods, leading to enhanced performance on a wide range of tasks.
https://allenai.github.io/lumos/