Skip to main content
All Collections

Astronomy

LLMs in dynamical (and general) astronomy

No spamNo marketingOnly data
ES

Evgeny Smirnov

10 papers · 2 Must Read · 2023–2026

Last updated Aug 2, 2026

Sorted by publication date, newest first. New papers are marked so you can spot recent additions.

Introduction

How large language models are transforming astronomical research in celestial mechanics and dynamical astronomy.

1
Must Read
intermediate
★ Essential

Evaluating multimodal commercial and open-source large language models for dynamical astronomy: a benchmark study of resonant behavior classification(pdf)

Smirnov, Evgeny, Carruba, Valerio · 2026 · Scientific Reports

At a GlanceAI

Benchmark shows multimodal LLMs can classify orbital resonances from images with high accuracy without fine-tuning.

SummaryAI

The study introduces reproducible benchmarks for testing multimodal LLMs on classifying mean-motion and secular resonances from plots of resonant arguments. Commercial models achieved perfect scores on unambiguous cases and up to 94% F1 on a harder three-class dataset, while open-source models approached commercial performance on the full binary benchmark. The main weakness was identifying transient and resonance-sticking behavior, but the results indicate that even untuned, locally runnable models can be useful for dynamical-astronomy classification.

Have no resources to use the commercial models? Use OSS! You can even launch them on your laptop. And they ARE good.

— ES

Method:AI
Benchmarking multimodal commercial and open-source LLMs on image-based classification of dynamical resonance behavior.
Background:AI
Basic knowledge of orbital dynamics, mean-motion and secular resonances, and machine-learning evaluation metrics.
2
Niche
intermediate

AstroMMBench: A Benchmark for Evaluating Multimodal Large Language Models Capabilities in Astronomy(pdf)

Shi, Jinghang, Tang, Xiaoyu, Huang, Yang et al. · 2025 · arXiv (Cornell University)

At a GlanceAI

AstroMMBench tests multimodal LLMs on 621 expert-reviewed astronomy image questions across six subfields.

SummaryAI

AstroMMBench addresses the gap between general-purpose multimodal AI benchmarks and the specialized visual reasoning needed in astronomy. It provides 621 expert-reviewed multiple-choice questions spanning six astrophysical subfields, then compares 25 open- and closed-source multimodal language models. Ovis2-34B achieved the highest reported overall accuracy of 70.5%, while large performance differences across subfields—especially cosmology and high-energy astrophysics—show where models remain unreliable. The benchmark offers a domain-specific resource for measuring and improving AI systems intended for astronomical research.

Method:AI
The authors build an expert-curated astronomy image benchmark and evaluate 25 multimodal large language models on it.
Background:AI
Basic familiarity with astronomy image interpretation and multimodal large language models is helpful.
3
Worth Reading
intermediate

Achieving GPT-4o level performance in astronomy with a specialized 8B-parameter large language model(pdf)

Haan, Tijmen de, Ting, Yuan-Sen, Ghosal, Tirthankar et al. · 2025 · Scientific Reports

At a GlanceAI

AstroSage, an 8B astronomy LLM, matches GPT-4o on AstroMLab-1 through domain-specific training.

SummaryAI

AstroSage-Llama-3.1-8B is a freely available language model specialized for astronomy, astrophysics, cosmology, and astronomical instrumentation. It was trained on astronomy arXiv papers, additional astronomical literature, and millions of synthetic question-answer pairs, and achieves 80.9% on the AstroMLab-1 benchmark. Its reported GPT-4o-level performance shows that targeted domain training can allow relatively small open models to rival much larger general-purpose systems, potentially broadening AI support for astronomy education and research.

An important conclusion and point from this research is that you don't need a frontier model to achieve acceptable results!

— ES

Method:AI
Domain-specializing an 8B-parameter Llama model using astronomy literature and synthetic question-answer data, then benchmarking it on AstroMLab-1.
Background:AI
Basic familiarity with large language models, scientific benchmarks, and astronomy research literature.
4
Niche
advanced
★ Essential

APBench and benchmarking large language model performance in fundamental astrodynamics problems for space engineering(pdf)

Wu, Di, Zhang, Raymond, Zucchelli, Enrico M. et al. · 2025 · Scientific Reports

At a Glance

How good can LLMs solve space science university-level problems

Summary

Authors created a dataset of questions from Astrodynamics, tested a variety of LLMs including open-source ones on them, and evaluated their performance. The paper is a good example of a benchmark study and how to conduct it. Helpful for anyone doing benchmark stuff in astronomy.

A nice example of the usefulness of LLMs in astronomy + a good example of how to do a benchmark study in astronomy + LLM

— ES

Method:
LLM (benchmark)
Background:
Deep knowledge of benchmarking in LLMs + the state of art
5
Skip
intermediate

Deep Learning and Methods Based on Large Language Models Applied to Stellar Light Curve Classification

Li, Yu-Yang, Bai, Yu, Wang, Cunshi et al. · 2025 · Intelligent Computing

At a GlanceAI

StarWhisper LC adapts language-model variants to classify stellar light curves at about 90% accuracy with little feature engineering.

SummaryAI

The study evaluates deep learning and LLM-based approaches for automatically classifying variable-star light curves from the Kepler and K2 missions. Its main novelty is the StarWhisper LC series, which fine-tunes language, multimodal-language, and audio-language models for astronomical time-series data and achieves around 90% accuracy while reducing explicit feature engineering. Conventional optimized models perform even better overall, with the Swin Transformer reaching 99% accuracy and identifying rare type II Cepheids at 83% accuracy. The accompanying analyses of cadence and phase coverage suggest that observations can be shortened or sampled less densely with limited loss of classification performance.

Method:AI
The study benchmarks optimized deep-learning architectures and fine-tuned LLM, multimodal LLM, and audio-language models for light-curve classification.
Background:AI
Basic machine learning and astronomy knowledge, especially variable stars and time-series light curves.
6
Worth Reading
beginner

At First Sight! Zero-shot Classification of Astronomical Images with Large Multimodal Models(pdf)

Tanoglidis, Dimitrios, Jain, Bhuvnesh · 2024 · Research Notes of the AAS

At a GlanceAI

GPT-4o and LLaVA-NeXT classify galaxy images and artifacts above 80% accuracy using prompts alone.

SummaryAI

The study tests whether large vision-language models can classify astronomical images without being trained on astronomy-specific labels. GPT-4o and the open-source LLaVA-NeXT achieve typically above 80% accuracy for low-surface-brightness galaxies, artifacts, and galaxy morphology using natural-language prompts. The results position multimodal LLMs as potentially useful research and teaching tools, while highlighting that open models such as LLaVA-NeXT still need improvement and may benefit from domain-specific fine-tuning.

Yep, that's the point: no training, no code, just use. Similar to Smirnov (2024).

— ES

Method:AI
Evaluating GPT-4o and LLaVA-NeXT with natural-language prompts for zero-shot classification of astronomical images.
Background:AI
Basic knowledge of galaxy morphology, astronomical imaging, and vision-language models is helpful.
7
Niche
intermediate

Designing an Evaluation Framework for Large Language Models in Astronomy Research(pdf)

Wu, John F., Hyk, Alina, McCormick, Kiera et al. · 2024 · arXiv (Cornell University)

At a GlanceAI

A proposed framework evaluates how astronomers use and assess an arXiv-grounded RAG chatbot in real research settings.

SummaryAI

The paper addresses the lack of a standard way to evaluate LLM tools specifically for astronomy research. It presents an experimental design built around a Slack chatbot whose answers are grounded in astronomy papers on arXiv, while recording anonymized questions, answers, retrieval evidence, ratings, and feedback. This setup supports ongoing, real-world evaluation of usefulness and failure modes, helping guide the development of more reliable LLM systems for astronomers.

Method:AI
Deploying a Slack-based RAG chatbot grounded in arXiv astronomy papers and collecting anonymized interaction and feedback data.
Background:AI
Basic knowledge of large language models, retrieval-augmented generation, and astronomy research literature.
8
Must Read
beginner
★ Essential

Fast, Simple, and Accurate Time Series Analysis with Large Language Models: An Example of Mean-motion Resonances Identification(pdf)

Smirnov, Evgeny A. · 2024 · The Astrophysical Journal

At a Glance

LLM can classify resonant behavior without coding or knowledge with high accuracy

Summary

The author created LLM pipeline with just one prompt that classifies resonant arguments and identifies whether time series has libration or circulation. This is a pilot study and simple cases are considered, but the accuracy and F1 score of 100% are remarkable. What's more important: instead of coding and using special methods, it takes 15m to write a prompt and use it.

Forget about filtering, periodograms, and working with time series. Just use LLMs!

— ES

Method:
LLM
Background:
Basic knowledge of LLMs
9
Niche
advanced

Compact binary systems waveform generation with a generative pretrained transformer

Shi, Ruijun, Zhou, Yue, Zhao, Tianyu et al. · 2024 · Physical Review D

At a GlanceAI

Generative pretrained transformers are applied to produce gravitational-wave waveforms from compact binary systems.

SummaryAI

This study brings generative pretrained transformer methods to the task of producing gravitational-wave signals from compact binary systems. It connects large-language-model-style generative architectures with a central computational problem in gravitational-wave astronomy: waveform generation. The work may broaden the machine-learning toolkit for modeling compact-binary signals and motivates further evaluation of transformer-based waveform models in scientific inference pipelines.

Method:AI
Applying a generative pretrained transformer to generate compact-binary gravitational-wave waveforms.
Background:AI
Gravitational-wave physics, compact-binary waveform modeling, and transformer-based machine learning.
10
Niche
beginner

Astronomia ex machina: a history, primer and outlook on neural networks in astronomy(pdf)

Smith, Michael J., Geach, James E. · 2023 · Royal Society Open Science

At a GlanceAI

Review advocating open, astronomy-specific GPT-like foundation models for multimodal astronomical data.

SummaryAI

This review charts three waves of neural networks in astronomy, from early multilayer perceptrons through convolutional and recurrent networks to unsupervised and generative deep learning. It argues that rapidly growing, multimodal astronomical datasets make GPT-like foundation models a promising next step for supporting many downstream astronomy tasks. The authors propose that the astronomy community collaboratively develop open-source foundation models rather than relying solely on systems driven by large technology companies.

Has some nice theoretical stuff.

— ES

Method:AI
A historical and forward-looking review of neural-network methods in astronomy, culminating in a proposal for astronomy-specific foundation models.
Background:AI
Basic knowledge of machine learning, neural networks, and astronomical data analysis is helpful.