An agentic framework for gravitational-wave counterpart association in the multi-messenger era

Summary

This paper presents GW-Eyes, an agentic framework powered by large language models (LLMs) designed for gravitational-wave (GW) counterpart association in the multi-messenger era. The authors state: "To help address these challenges, we develop GW-Eyes, an agentic framework powered by large language models (LLMs). For the first time, GW-Eyes integrates domain-specific tools and autonomously performs counterpart association tasks between GW and candidate EM events."

The motivation stems from the fact that With the detection of gravitational waves (GWs), multi-messenger astronomy has opened a new window for advancing our understanding of astrophysics, dense matter, gravitation, and cosmology. The paper notes that GW170817 was the first binary neutron star (BNS) merger detected through gravitational waves (GWs) and accompanied by multi-wavelength electromagnetic (EM) counterparts, and that "Joint observations of the GW signal, the associated short-duration gamma-ray burst (SGRB) GRB 170817A, and the kilonova transient AT 2017gfo, have yielded remarkable scientific returns, including a smoking-gun evidence for the BNS merger origin of SGRBs and kilonovae, and an independent measurement of the Hubble constant via the standard siren method. The authors also note that EM counterparts are also expected from neutron star–black hole (NS–BH) mergers, as well as stellar-mass binary black hole (BBH) mergers in active galactic nucleus (AGN) disks, but GW follow-up campaigns have so far been unsuccessful to identify convincing EM counterparts to NS–BH or BBH mergers."

The framework architecture is described as follows: "GW-Eyes is an LLM-driven agentic system designed for GW counterpart association. As illustrated in Figure 1a, it comprises two sub-agents, Collector and Executor, and integrates domain-specific tools through the Model Context Protocol. Collector is responsible for data acquisition and initial preprocessing, including the retrieval of GW posterior information, public EM event catalogs, and real-time circulars from General Coordinates Network (GCN). Executor operates on the database to perform counterpart association tasks, as well as auxiliary functions such as visualization and rapid verification."

The paper details the workflow: "Upon receiving a data-acquisition task, such as updating GCN circular records, Collector first retrieves the relevant online GCN circular contents... Collector then automatically interprets and extracts the information relevant to counterpart association, storing it in the database according to predefined rules and thereby enabling subsequent analyses by the Executor. For counterpart association tasks, Executor first decomposes the request and formulates an analysis strategy. It then invokes tools to access GW posteriors and EM catalogs, and evaluates their spatial associations using statistical metrics. After generating a list of candidates, Executor can provide visualizations, including sky-localization maps and the luminosity distance posterior, together with a detailed report summarizing the selection criteria and final results. The association analysis can be performed in either direction: starting from a GW event to identify potential EM counterparts, or from an EM transient to identify candidate GW events."

The authors emphasize the advantages over general-purpose LLMs: "the agentic design of GW-Eyes and its integration of domain-specific tools provide clear advantages over general-purpose commercial LLMs. For instance, GW-Eyes can build and maintain a local event database with consecutive updates, while supporting secure access to private data. In the highly specialized multi-messenger astronomy, general-purpose LLMs often struggle to process domain-specific data effectively, whereas our integration of dedicated tools substantially improves the adaptability of GW-Eyes."

For evaluation, the authors constructed a simulated EM catalog: "We select the first 50 GW events from GWTC-4.0, that span from May 18 to October 4, 2023. For each GW event, we inject one ground-true EM counterpart occurring within 7 days after the GW trigger, located inside the 50% credible region of the GW localization, and with a redshift in the 68% interval of its distribution. We then add two confounding events that satisfy the temporal and sky-position criteria but with a false redshift, and ten unrelated events that satisfy only the temporal criterion. The simulated catalog contains 650 EM events and is used to test whether GW-Eyes correctly recover the injection."

The evaluation covered two tasks: GW4EM, which retrieves the injected EM counterpart for a given GW event, and EM4GW, which retrieves the associated GW event for a given EM event. The results show that GW-Eyes consistently achieves stable counterpart recovery across different LLM backends (GPT-5, Gemini-3-flash-preview, and claude-sonnet-4-6), while maintaining a relatively low per-task cost. The paper reports recall rates of 100% for all models on both tasks, with input/output token counts and costs detailed in Figure 1c.

The authors also demonstrated application to real catalogs: "Given that most events in GWTC-4.0 are BBH mergers, we take AGN optical flares as an example to explore potential EM counterpart association. We use the Active Galactic Nuclei Flare Refined Catalog (AGNFRC), whose temporal coverage overlaps with GWTC-4.0. We adopt a temporal window of 180 days after each GW trigger, apply GW-Eyes to search for potential temporal and spatial associations between events in the two catalogs, and report candidate counterparts whose sky-position falls within the 50% and 90% credible regions of the GW localization. For events with reported redshifts, we further require consistency within the 68% credible region. The paper notes that Given the relatively long time window adopted and the relatively poor sky localization of some GW events, temporal and spatial associations alone can yield more than one EM candidate for certain GWs."

The authors computed "the number of associated GW-EM pairs (NGW−EM) with different GW localization area thresholds, normalized by the number of EM events (NEM) within the adopted 180-day post-trigger temporal windows of the 50 GWTC-4.0 events, and show the cumulative ratio NGW−EM/NEM as a function of the GW skymap area threshold. They conclude that Given that in the future we are likely to have O(10 5) EM transient alerts per night, Figure 1e further shows that counterpart association will face substantial challenges when the numbers of GW and EM events increase rapidly. Fortunately, GW-Eyes can substantially help address these challenges and support different groups in tailoring its use to specific needs, greatly facilitating subsequent identification and follow-up studies."

The paper highlights key advantages: "Compared to other machine-learning-based approaches, GW-Eyes, built on an LLM agent framework, provides more interpretable reasoning traces and results. On the one hand, the full decision-making process, including tool usage and parameters, is recorded in the human-readable natural language, facilitating inspection and verification. On the other hand, compared with general-purpose LLMs, the integration of domain-specific tools significantly reduces hallucinations in the specialized context of multi-messenger astronomy, therefore improving the accuracy and reliability of its results."

The authors emphasize extensibility: "GW-Eyes is highly extensible. In the present study, counterpart association relies primarily on basic temporal and spatial overlap between GW and EM signals, while more specialized criteria can be readily incorporated by extending the agent's toolset. For different EM datasets, this may include additional information such as spectra and images to support refined association strategies. The advanced decision-making capabilities of LLMs enable flexible workflows beyond the simple if-else logic, offering strong potential for extensions."

The conclusion states: "This Letter presents GW-Eyes, a first LLM-powered agent framework for GW counterpart association in multi-messenger astronomy. Through the design of Collector and Executor sub-agents and the integration of domain-specific tools, GW-Eyes enables EM catalog management, counterpart association analyses, and auxiliary support for human experts. Natural-language interaction significantly improves flexibility and provides interpretable reasoning traces that facilitate subsequent inspection and verification. In addition, GW-Eyes is highly extensible, allowing more specialized association strategies to be readily incorporated for different EM catalogs. We expect GW-Eyes to offer a promising approach to alleviate the growing human workload in the big-data era of multi-messenger astronomy."

The Supplementary Information provides three examples of GW-Eyes usage: counterpart-association tasks for BBH mergers (Figure A1), catalog updates based on GCN Circulars (Figure A2), and rapid verification tasks (Figure A3). These examples demonstrate the tool-calling procedures and outputs, including detailed reasoning traces, candidate identification, and visualizations.

Based on the paper, here are specific improvements I can make to AI systems:

Improvement: Implement a Collector-Executor dual-agent design where one agent handles data acquisition/parsing (e.g., extracting structured data from GCN circulars) and another handles analysis/decision-making. This separation improves modularity and error isolation.

Capability: The system can autonomously ingest unstructured astronomical text (e.g., GCN notices), convert it to structured database entries, and then perform complex queries—all without human intervention.

Improvement: Integrate domain-specific tools (skymap querying, coordinate matching, distance validation) directly into the LLM's tool-calling loop, rather than relying on the model's parametric knowledge.

Improvement: Design the reasoning engine to support reverse lookup—starting from an EM event and finding candidate GW events—not just forward association.

Improvement: Implement a hierarchical filtering pipeline that applies temporal windows (e.g., ±180 days), spatial containment (50%/90% credible regions), and redshift/distance consistency (within 1σ) sequentially.

Improvement: Force the LLM to log every tool call, parameter choice, and intermediate reasoning step in natural language, creating a complete audit trail.

Improvement: Implement logic to automatically select the most appropriate waveform model's skymap (e.g., IMRPhenomXPHM vs. NRSur7dq4) based on event characteristics, with fallback options.

Improvement: Add a dedicated sub-agent that periodically fetches new GCN circulars, parses them, and updates the local database without user prompting.

Improvement: Implement a minimal-context strategy where only essential data (e.g., skymap file paths, not full FITS files) are passed to the LLM, with heavy computations delegated to tools.

Improvement: Design the tool interface to accept pluggable validation functions (e.g., spectral matching, image-based classification) beyond basic spatial/temporal filters.

Improvement: Integrate plotting tools that generate skymap overlays and distance posterior distributions, then automatically embed them into a structured markdown report.