Graph clustering aims to group nodes into meaningful clusters, a crucial task for understanding the behavior of complex networks and their underlying structural organization. Leveraging the expressive power of deep learning, graph neural networks (GNNs) have emerged as popular tools for performing graph clustering with high accuracy. However, despite their promise, existing GNN-based approaches often lack structural reasoning mechanisms to adaptively infer and optimize each node's receptive field in a task-specific manner, limiting their potential to further improve both accuracy and interpretability. In this paper, we propose a novel reinforced structural reasoning framework, termed RGIGC, for interpretable graph clustering. RGIGC formulates receptive field configuration as a structural reasoning problem and employs reinforcement learning with two Q-learning agents to infer and optimize the structural receptive field of each node. These agents are guided by reward criteria that strategically shape the clustering process to achieve desired outcomes. By identifying critical structures of receptive fields that capture the intrinsic features of nodes, RGIGC can enhance the interpretability of the resulting clusters. Extensive experiments on eight real-world graph datasets of varying scales have been conducted, and the results demonstrate that RGIGC outperforms state-of-the-art graph clustering algorithms in both accuracy and interpretability.
Yue Yang, Dongxu Li, Hengchuang Yin et al.· Proceedings of the 32nd ACM...· 0 citations
The optimization of long-horizon agents increasingly relies on reflection-based mechanisms, where a large language model (LLM) acts as an optimizer to diagnose agent failures and improve agent policies. However, real execution traces are difficult to use directly for optimization: large trace collections are often redundant and heterogeneous, making optimization inefficient and prone to overfitting to low-value failures; meanwhile, each individual trajectory also contains many irrelevant steps, while naive context reduction methods such as truncation or sliding windows can discard causally important evidence and produce misleading optimization signals. To resolve this dilemma, we introduce STRACE (Structural TRajectory Analysis and Causal Extraction), a framework that constructs high signal-noise optimization contexts for more precise and effective optimization. At the batch level, STRACE mines failure patterns to filter redundant traces and retain representative failures; within each selected trace, it performs causal localization over a textual dependency graph to remove non-causal steps and identify the true root-cause module for optimization. Empirical results demonstrate that STRACE significantly outperforms standard context-filtering baselines. Notably, on a challenging formal verification task (VeruSAGE-Bench), it successfully optimizes human-expert designed agents, delivering $1.4\times$ success-rate improvement (42.5% to 58.5%). The code is available at https://github.com/moomight/STRACE .
Ying Chang, Jiahang Xu, Xuan Feng et al.· 0 citations