Generative Flow Networks (GFlowNets) are a powerful tool for generating structured objects with diverse, high-reward outcomes by sampling from a distribution proportional to a given reward function. Unlike traditional reinforcement learning (RL) approaches, GFlowNets aim to balance diversity and reward by modeling the entire trajectory distribution. This makes them suitable for domains such as molecular design and combinatorial optimization. However, existing GFlowNets sampling strategies often lead to excessive exploration and struggle to consistently generate high-reward samples, especially in large exploration spaces with sparse high-reward regions. In this study, we integrate an enhanced Monte Carlo Tree Search (MCTS) into the GFlowNets sampling process, inducing the generation of high-reward trajectories through MCTS-based policy evaluation. We adaptively balance exploration and exploitation using Polynomial Upper Confidence Trees (PUCT), and introduce a controllable greedy mechanism. Our method dynamically balances exploration and reward-based guidance without sacrificing diversity, thereby enhancing exploitation.
