Bowen Zhou
Publications
MARS$^2$: Scaling Multi-Agent Tree Search via Reinforcement Learning for Code Generation
Reinforcement learning (RL) paradigms have demonstrated strong performance on reasoning-intensive tasks such as code generation. However, limited trajectory diversity often leads to diminishing returns, which constrains the achievable performance ceiling. Search-enhanced RL alleviates this issue by introducing structured exploration, which remains constrained by the single-agent policy priors. Meanwhile, leveraging multiple interacting policies can acquire more diverse exploratory signals, but existing approaches are typically decoupled from structured search. We propose \textbf{MARS$^2$} (Multi-Agent Reinforced Tree-Search Scaling), a unified RL framework in which multiple independently-optimized agents collaborate within a shared tree-structured search environment. MARS$^2$ models the search tree as a learnable multi-agent interaction environment, enabling heterogeneous agents to collaboratively generate and refine candidate solutions within a shared search topology. To support effective learning, we introduce a path-level group advantage formulation based on tree-consistent reward shaping, which facilitates effective credit assignment across complex search trajectories. Experiments on code generation benchmarks show that MARS$^2$ consistently improves performance across diverse model combinations and training settings, demonstrating the effectiveness of coupling multi-agent collaboration with tree search for enhancing reinforcement learning. Our code is publicly available at https://github.com/TsinghuaC3I/MARTI.
Empowering Private Tutoring by Chaining Large Language Models
Artificial intelligence has been applied in various aspects of online education to facilitate teaching and learning. However, few approaches has been made toward a complete AI-powered tutoring system. In this work, we explore the development of a full-fledged intelligent tutoring system powered by state-of-the-art large language models (LLMs), covering automatic course planning and adjusting, tailored instruction, and flexible quiz evaluation. To make the system robust to prolonged interaction and cater to individualized education, the system is decomposed into three inter-connected core processes-interaction, reflection, and reaction. Each process is implemented by chaining LLM-powered tools along with dynamically updated memory modules. Tools are LLMs prompted to execute one specific task at a time, while memories are data storage that gets updated during education process. Statistical results from learning logs demonstrate the effectiveness and mechanism of each tool usage. Subjective feedback from human users reveal the usability of each function, and comparison with ablation systems further testify the benefits of the designed processes in long-term interaction.