A selection of papers, announcements and tools on generative AI in astrophysics and scientific research, picked from what circulates in the Gen.IA community, most recent first. Open each item for the summary and the link to the source.
Verification by a group of scientists, including a Nobel laureate, of the result OpenAI published in October on the $1 million Yang-Mills quantum theory problem.
Anthropic reports that two black hole problems were solved, with both papers written entirely by Claude Fable 5.2. No details are given on how the result was obtained or which prompts were used.
CERN and Mistral’s plan for a foundation model for LHC reconstruction, plus the Large Physics Models paper and the European Coalition for AI in Fundamental Physics.
In response to a public challenge, two Anthropic physicists had Claude compute the six-particle amplitude in planar N=4 super Yang-Mills at nine loops, one beyond the previous record. The run was largely unsupervised.
Anthropic’s interpretability team describes a small set of internal representations in Claude that behaves like the global workspace of cognitive science. The authors make no claim about consciousness.
A system combining Gemini Deep Think with tree search and numerical checks derived exact analytical solutions for the spectrum of gravitational radiation emitted by cosmic strings.
A system that proposes ideas, runs experiments, writes the manuscript and reviews it. One fully AI-generated paper passed peer review for a conference workshop. For now it only handles computational experiments.
A study of whether journals’ AI policies have changed how much AI-assisted writing appears in academic papers. The authors find that the policies have not curbed its growth.
Case studies in which researchers worked with Gemini Deep Think on open problems in theoretical computer science, economics, optimisation and physics, with the collaboration techniques that kept recurring.
A multi-agent system built as a research assistant: it generates ideas, checks the literature, writes and runs code, and drafts papers. The authors report expert evaluations and discuss its limitations.
We present LEMUR, a publicly accessible archive of uniformly reduced, deep Chandra images of 2,569 galaxies and galaxy clusters (including groups), comprising 5,505 ACIS observations and 138.8 Ms of exposure. By combining the publicly available imaging observations of…
The search for better models, whether expressed as analytical equations or neural architectures, has increasingly shifted from manual design toward data-driven automation. While a multitude of algorithms exist, ranging from exhaustive search to AlphaEvolve, recent advances in general-purpose…
The hydroxyl radical (OH) is the main oxidant in the troposphere and accounts for 88% of the methane sink, yet its abundance and trends remain poorly constrained. Satellites observe many of the chemical and meteorological drivers of OH,…
In cosmology, surveys span different sky regions and wavelength ranges, resulting in uneven modality coverage across observational datasets. Such heterogeneity can be addressed through either multimodal foundation models (FMs), which learn shared embedding spaces, or independently learned modality…
Spectra from different surveys carry different systematics and are not calibrated against one another. Multi-survey foundation models for stellar spectra are then built by aligning embeddings of the same star observed across surveys. Most in astronomy follow the…
New software tools have emerged that enable AI agents to quickly develop code that analyzes data and generates figures, but their mistakes can also harm a researcher's computer, require extensive resources for validation, and cause the results to…
Galactic Chemical Evolution (GCE) models connect stellar nucleosynthesis to observed abundance patterns, but their interpretation requires simultaneous parameter inference and comparison between competing stellar yield prescriptions. This remains challenging due to intractable likelihoods and growing dataset complexity. Using…
Lagrangian particle tracking follows virtual particles in the velocity fields of an ocean model to study transport pathways, connectivity and dispersion. Setting up such a study takes many technical steps, and this work is increasingly done with coding…
Selected automatically from arXiv astro-ph by keyword (LLMs, generative AI, foundation models, agents, diffusion models) and refreshed every few hours. Not reviewed by Gen.IA. Thank you to arXiv for use of its open access interoperability.
From the community
Shared on the Gen.IA blog
Links and notes posted by Gen.IA members, most recent first.
Physicist Nathaniel Craig (UC Santa Barbara) describes how OpenAI’s GPT-6 Astra took the lead in the detailed proof of a candidate lattice definition of the Standard Model, a problem long blocked by the difficulty of putting chiral fermions on a lattice. The preprint (arXiv:2610.09093) has not yet been peer reviewed.
Verification by a group of scientists, including a Nobel laureate, of the result OpenAI published in October on the $1 million Yang-Mills quantum theory problem.
Anthropic reports that two black hole problems were solved, with both papers written entirely by Claude Fable 5.2. No details are given on how the result was obtained or which prompts were used.
CERN and Mistral’s plan for a foundation model for LHC reconstruction, plus the Large Physics Models paper and the European Coalition for AI in Fundamental Physics.
An LLM-driven toolkit for exact quantitative science: project site, arXiv paper and a related Anthropic post.
Ask Gen.IA
Ask about Gen.IA, its activities, resources and highlights. Answers are generated by AI from the content of this website and may contain mistakes: check the linked pages.
Questions are sent to an external AI service through OpenRouter. Do not include personal or confidential data.