AdaSTORM: Scaling LLM Reasoning on Dynamic Graphs via Adaptive Spatio-Temporal Multi-Agent Collaboration

A new framework called AdaSTORM enables large language models to reason over dynamic graphs containing thousands of nodes, a scale far beyond the tens of nodes that current models can handle, according to research posted to the arXiv preprint server on June 15 [1][2]. Large language models, or LLMs, are machine learning models trained on vast text corpora for tasks such as language generation [8]. Their application to dynamic graph reasoning has been constrained by exponential reasoning overhead and finite context windows, limiting them to graphs with tens of nodes [2]. The new paper introduces AdaSTORM, which stands for Scaling LLM Reasoning on Dynamic Graphs via Adaptive Spatio-Temporal Multi-Agent Collaboration [1][2]. The framework reformulates large-scale dynamic graph reasoning into two stages. The first, Adaptive Partitioning, divides large-scale dynamic graphs into subregions matched to the model’s reasoning capacity while minimizing inference cost. The second, Collaborative Reasoning, aligns graph partition topologies with a spatio-temporal decoupled multi-agent architecture [1][2]. The authors describe AdaSTORM as the first multi-agent framework tailored for dynamic graph reasoning [2]. In experiments, AdaSTORM scaled reasoning to thousand-node graphs with over 90% accuracy across several large-scale dynamic graph settings, without relying on external tools [1][2]. It outperformed seven competitive baselines, achieved state-of-the-art accuracy on existing benchmarks, and generalized robustly to real-world datasets [1][2]. The source code has been made available on GitHub [2]. The paper was submitted to arXiv’s artificial intelligence section. arXiv is an open-access repository of electronic preprints, founded in 1991, that hosts scientific papers across mathematics, physics, computer science, and related fields [6]. As of November 2024, the repository receives about 24,000 new articles per month [6]. Papers on arXiv are moderated but not peer-reviewed [6]. The AdaSTORM preprint appears alongside a suite of experimental community tools offered through arXivLabs, a framework that allows collaborators to develop and share features such as bibliographic explorers and code finders directly on the abstract page [4][5].