标题： 通过推理时树搜索引导的扩散模型进行可控制图生成 显示英文标题

标题： Controllable Graph Generation with Diffusion Models via Inference-Time Tree Search Guidance

摘要： 图生成是图学习中的基础问题，在Web规模系统、知识图谱以及药物和材料发现等科学领域有广泛的应用。 最近的方法利用扩散模型进行逐步生成，但无条件扩散对所需属性控制有限，通常会导致质量不稳定，并且在引入新目标时存在困难。 推理时引导方法通过调整采样过程而不重新训练来缓解这些问题，但它们本质上是局部的、启发式的，并且可控性有限。 为了克服这些限制，我们提出了TreeDiff，这是一种蒙特卡洛树搜索（MCTS）引导的双空间扩散框架，用于可控制的图生成。 TreeDiff是一种即插即用的推理时方法，在保持计算可行性的前提下扩展了搜索空间。 具体来说，TreeDiff引入了三个关键设计，使其具有实用性和可扩展性：(1) 宏步骤扩展策略，将多个去噪更新分组为一个转换，减少树的深度并实现长视界探索；(2) 双空间去噪机制，将高效的潜在空间去噪与图空间中的轻量离散修正相结合，确保可扩展性和结构保真度；以及(3) 双空间验证器，从部分去噪的图中预测长期奖励，实现早期价值估计并消除对完整回放的需求。 在2D和3D分子生成基准测试中，无论是在无条件还是条件设置下，广泛的实验都表明TreeDiff实现了最先进的性能。 值得注意的是，TreeDiff在推理时具有良好的扩展性：随着计算资源的增加，它持续改进，而现有的推理时方法在资源有限的情况下会早期达到瓶颈。 显示英文摘要

摘要： Graph generation is a fundamental problem in graph learning with broad applications across Web-scale systems, knowledge graphs, and scientific domains such as drug and material discovery. Recent approaches leverage diffusion models for step-by-step generation, yet unconditional diffusion offers little control over desired properties, often leading to unstable quality and difficulty in incorporating new objectives. Inference-time guidance methods mitigate these issues by adjusting the sampling process without retraining, but they remain inherently local, heuristic, and limited in controllability. To overcome these limitations, we propose TreeDiff, a Monte Carlo Tree Search (MCTS) guided dual-space diffusion framework for controllable graph generation. TreeDiff is a plug-and-play inference-time method that expands the search space while keeping computation tractable. Specifically, TreeDiff introduces three key designs to make it practical and scalable: (1) a macro-step expansion strategy that groups multiple denoising updates into a single transition, reducing tree depth and enabling long-horizon exploration; (2) a dual-space denoising mechanism that couples efficient latent-space denoising with lightweight discrete correction in graph space, ensuring both scalability and structural fidelity; and (3) a dual-space verifier that predicts long-term rewards from partially denoised graphs, enabling early value estimation and removing the need for full rollouts. Extensive experiments on 2D and 3D molecular generation benchmarks, under both unconditional and conditional settings, demonstrate that TreeDiff achieves state-of-the-art performance. Notably, TreeDiff exhibits favorable inference-time scaling: it continues to improve with additional computation, while existing inference-time methods plateau early under limited resources.