Large Language Model Driven Analysis of General Coordinates Network (GCN) Circulars

The Big Picture

Imagine a fire department that responds to every emergency in a city, but instead of radio dispatches, each incident report arrives as a handwritten note, in slightly different formats, from thousands of different writers, accumulated over thirty years. That’s roughly the situation astronomers face with the General Coordinates Network (GCN), NASA’s global alert system for astronomical explosions, colliding neutron stars, and other cosmic catastrophes.

Every time something spectacular happens in the universe, observatories around the world fire off rapid reports called GCN Circulars: human-written, free-form dispatches describing what was seen, in what wavelengths, from where. A gamma-ray burst (an extraordinarily powerful explosion briefly brighter than an entire galaxy), a gravitational wave event, a neutrino flash from a distant source. Over three decades, more than 40,500 of these Circulars have piled up into a rich but unwieldy scientific archive.

The problem? They’re nearly impossible to search systematically. No standard format, no mandatory fields, no consistent vocabulary. A scientist who wants to find every gamma-ray burst with a measured redshift (a number encoding how far away an explosion occurred, derived from the stretching of light by cosmic expansion) has to dig through the pile by hand.

A team led by researchers at NASA Goddard and IAIFI has now shown that large language models (AI systems trained on vast text corpora to understand and reason about language) can crack this problem, automating classification, topic organization, and precise data extraction with near-human accuracy.

Key Insight: By combining neural topic modeling, contrastive fine-tuning, and retrieval-augmented generation, this system extracts gamma-ray burst redshifts from unstructured astronomical reports with 97.2% accuracy, all without task-specific training.

How It Works

The team built a multi-stage pipeline, each stage tackling a different layer of the Circulars’ messiness.

Stage 1: Neural topic modeling. First, they needed to understand what the archive is actually about. They used BERTopic, an embedding-based clustering tool that groups documents by meaning rather than keywords, to automatically discover the major themes across all 40,500 Circulars. Out fell clean clusters corresponding to gamma-ray bursts, gravitational wave events, X-ray transients, fast radio bursts, and their electromagnetic follow-up campaigns.

Stage 2: Waveband classification. Knowing a Circular’s topic isn’t enough; you also need to know how the event was observed. X-ray? Optical? Radio? The team used contrastive fine-tuning (a technique that trains a model by pushing similar examples closer together in embedding space and dissimilar ones apart) to classify Circulars by observation waveband and messenger type, including gravitational waves. This step made it possible, for the first time, to programmatically separate gravitational wave event clusters from the light-based follow-up observations that accompany a gravitational wave detection.

Stage 3: Redshift extraction. Redshift values are buried in prose: “The VLT spectrum shows an absorption feature at z = 1.24,” or “consistent with a host galaxy at redshift 0.35.” To pull them out, the team deployed Mistral, an open-source large language model, backed by three supporting techniques:

• Prompt tuning: carefully engineered instructions specifying exactly what format to extract data in
• Output parsing: structured post-processing that catches and corrects garbled or incomplete model responses
• Retrieval-Augmented Generation (RAG): rather than feeding every Circular to the model, the system first retrieves only the most relevant documents via a neural search index, then passes those to Mistral for extraction

None of this requires labeled training data. No annotated examples of “here is a Circular, here is the redshift.” The model generalizes entirely from prompt instructions and retrieval context.

Results were benchmarked against the manually curated Neil Gehrels Swift Observatory gamma-ray burst (GRB) table, a gold-standard reference built by human experts over years. The LLM pipeline correctly identified redshift values in 97.2% of Circulars that contained them, and the neural RAG retrieval correctly flagged 96.8% of redshift-containing documents from the broader archive.

Why It Matters

Astronomy is drowning in data, and the flood is only getting worse. The Rubin Observatory will generate millions of transient alerts per night. New gravitational wave detectors will multiply neutron star merger discoveries. Fast radio burst surveys are expanding rapidly. The GCN Circulars database will only grow faster, and without automation, the gap between data generated and data actually analyzed will keep widening.

The paper also makes a point worth paying attention to: LLMs can act as flexible, general-purpose scientific data extractors without retraining, as long as the retrieval and prompting infrastructure around them is well-designed. That’s a real proof of concept for astroinformatics, the growing effort to use AI to mine the accumulated written record of astronomy. The same approach could extend to other scientific alert networks, arXiv preprints, or observation logs buried in legacy databases.

Every field has its pile of unstructured text. This team just showed how to start reading it.

Bottom Line: A surprisingly simple combination of off-the-shelf LLMs and retrieval tools can extract structured data from 30 years of messy astronomical reports with near-perfect accuracy, pointing toward how AI can tap into science’s vast unstructured archives.