TreeMind : Automatically Reproducing Android Bug Reports via LLM-empowered Monte Carlo Tree Search

Automatically reproducing Android app crashes from textual bug reports is challenging, particularly when the reports are incomplete and modern UIs exhibit high combinatorial complexity. Approaches relying on reinforcement learning or large language models (LLMs) struggle to infer missing steps and reconstruct user action sequences in vast UI interaction spaces. We present TreeMind, which integrates LLMs with an adapted Monte Carlo Tree Search (MCTS) for strategic UI exploration in bug reproduction. TreeMind formulates the reproduction as a target-driven search problem and leverages MCTS to iteratively refine action sequences. To enhance MCTS with semantic reasoning, we introduce two LLM-guided agents: Expander generates promising actions based on the current UI state and exploration history, while Simulator estimates the likelihood that each candidate action leads toward successful reproduction using dynamic environment feedback. With multi-modal UI inputs and tailored prompts, TreeMind performs feedback-aware navigation that identifies essential user actions and incrementally reconstructs reproduction paths. We evaluate TreeMind on 100 Android bug reports, including 90 from three benchmarks and 10 from GitHub. Experimental results show that it outperforms four state-of-the-art baselines in reproduction success rate and average time cost over all runs, suggesting the promise of integrating LLM reasoning with MCTS-based planning for automated bug reproduction.