An agentic framework for gravitational-wave counterpart association in the multi-messenger era
Dear Editor,
As a landmark discovery in astronomy and fundamental physics, GW170817 was the first binary neutron star (BNS) merger detected through gravitational waves (GWs) and accompanied by multi-wavelength electromagnetic (EM) counterparts.1 Joint observations of the GW signal, the associated short-duration gamma-ray burst (SGRB), and the kilonova transient have yielded remarkable scientific returns,1 including smoking-gun evidence for the BNS merger origin of SGRBs and kilonovae,2 and an independent measurement of the Hubble constant via the standard siren method.3 Besides BNSs, neutron star-black hole (NS-BH) mergers, as well as stellar-mass binary BH (BBH) mergers in active galactic nucleus (AGN) disks, have also been speculated to give rise to EM counterparts.4 Unfortunately, GW follow-up campaigns have so far been unsuccessful in identifying convincing EM counterparts to NS-BH or BBH mergers. The non-detections may result from the large distances of these events, incomplete coverage of their large GW localization regions, or intrinsically fainter EM emissions from them.
With the advent of next-generation GW detectors and advancing EM facilities, multi-messenger astronomy is expected to boost with richer scientific outcomes for all types of compact binary mergers discussed above. These opportunities, however, are accompanied by greater data-analysis challenges. In particular, searching for an association between GW signals and their EM counterparts is vital for enabling subsequent multi-messenger studies. The expected increase in co-detecting them will place growing pressure on both computational resources and human effort, preventing us from fully exploiting the scientific value of data.
The rapid advancement of artificial intelligence provides a promising pathway to help address emerging challenges. Machine learning has enabled real-time parameter estimation for BNS merger signals, delivering accurate estimates of sky localization within seconds,5 thereby supporting early alerts for multi-messenger follow-up. It has also been applied to searches for EM counterparts, assisting candidate prioritization in complex observational environments.6 Beyond conventional machine-learning approaches, the recent rise of large language models (LLMs) further shows the broad potential of artificial intelligence for scientific discoveries. Owing to their rapidly advancing reasoning capabilities, LLMs have been successfully explored in diverse scientific tasks, including automated scientific discovery.7 These LLM systems are able to achieve robust performance while substantially reducing the human effort and alleviating the pressure imposed by the rapidly growing data volume.
Motivated by this, we develop GW-Eyes, a first LLM-powered agentic framework designed to automate GW counterpart association tasks while assisting human experts through natural-language interaction with auxiliary functions such as catalog management, skymap visualization, and rapid verification. GW-Eyes integrates domain-specific tools tailored to the distinctive data structures of multi-messenger observations and dynamically retrieves and updates the latest information from the internet. We construct simulated datasets to evaluate the accuracy and efficiency of its tool use and further demonstrate its application to EM counterpart searches using real observational catalogs. Through GW-Eyes, we demonstrate how LLMs can participate in future multi-messenger observations and help address challenges in the big-data era.
