Searching for associations between gravitational-wave (GW) signals and their electromagnetic (EM) counterparts is essential for enabling multi-messenger studies and unlocking their rich scientific potential. It was demonstrated with GW170817, the first multi-messenger event from a binary neutron star merger, in establishing the origin of short gamma-ray burst, testing gravity theories, and measuring the expansion of the Universe. However, with the advent of next-generation GW detectors and continued advances in EM observing facilities, the expected rise in the number of joint detections will place growing demands on both computational resources and human effort, potentially limiting our ability to fully exploit the scientific value of these data. To address this challenge, a research team led by Prof. Lijing Shao at the Kavli Institute for Astronomy and Astrophysics, Peking University, has developed GW-Eyes, a framework that uses a large language model (LLM)-powered agent to automate GW–EM counterpart association, exploring how LLMs could contribute to future multi-messenger observations in the era of big data.
Figure 1: Schematic diagram of the GW-Eyes agent architecture.
An LLM agent extends a language model from a passive generator of responses into an active collaborator capable of pursuing complex goals. By perceiving its environment through data and tool interfaces, maintaining context, planning over multiple steps, and executing actions based on intermediate results, an agent can dynamically adapt its strategy as a task unfolds. This combination of language-based reasoning, tool use, and long-horizon decision-making makes LLM agents particularly promising for scientific workflows that require coordinating heterogeneous information and repeatedly deciding what to do next.
GW-Eyes combines the task-understanding and reasoning capabilities of LLMs with domain-specific tools. When presented with a scientific question, it can autonomously organize the analysis workflow and invoke appropriate tools for data processing and computation. Architecturally, GW-Eyes consists of two sub-agents: a Collector and an Executor. The Collector gathers multi-messenger astronomy data from multiple sources, while the Executor develops an analysis strategy based on the scientific objective and calls specialized tools to perform tasks such as catalog queries, spatial- and distance-consistency checks, candidate filtering, and visualization. GW-Eyes then produces an analysis report for researchers to inspect and verify. This design moves the role of LLMs beyond information generation toward the organization and orchestration of scientific workflows, offering a new approach to complex multi-messenger data analysis.
Figure 2: Illustration of the GW-Eyes analysis workflow.
To test whether the framework can reliably perform such association tasks, the research team constructed a simulated EM catalog based on 50 events from a GW catalog, GWTC-4.0, and designed two types of tasks: searching for EM counterparts of GW events and tracing GW events from EM transients. GW-Eyes consistently completed the prescribed tasks across different LLM backends. The team further applied GW-Eyes to real observational catalogs, searching for potential GW–EM associations between GWTC-4.0 and the active galactic nucleus (AGN) optical flare catalog, AGNFRC, using temporal, sky-position, and redshift information. An independent manual cross-check using the same analysis tools yielded 100% agreement, demonstrating the accuracy and robustness of GW-Eyes in counterpart-association tasks.
GW-Eyes provides the first demonstration of the potential of LLM agents for associating GW events with their EM counterparts, while also offering a new pathway for integrating LLM-based reasoning with domain-specific tools and specialized scientific data-analysis workflows.
The current work focuses primarily on information retrieval, candidate screening, association analysis, and result validation. As a next step, the GW-Eyes team plans to collaborate with high-energy transient follow-up teams and work with real observational data and analysis pipelines, starting from representative small-sample cases. The goal is to advance the agent toward end-to-end analysis in real-world observing scenarios, while further developing its capabilities for autonomous planning, dynamic decision-making, and workflow control in complex scientific tasks.
The study, “An agentic framework for gravitational-wave counterpart association in the multi-messenger era,” was published online in The Innovation. Yiming Dong, a PhD student in the Department of Astronomy, School of Physics at Peking University, is the first author. Prof. Lijing Shao is the corresponding author. Yacheng Kang and Ziming Wang, both PhD students in the Department of Astronomy at Peking University, associate researcher Junjie Zhao of the Henan Academy of Sciences, and Xinyuan Zhu from the University of Science and Technology of China, are co-authors.
Article: Y. Dong, Y. Kang, J. Zhao, X. Zhu, Z. Wang, L. Shao, The Innovation 7 (2026) 101538
Link:https://doi.org/10.1016/j.xinn.2026.101538
