What AstroPT knows about galaxies, and what that can teach us about LLMs

Hugging Face Daily Papers Papers
astronomy transformer-model interpretability linear-probes concept-emergence machine-learning

Summary

This paper proposes using AstroPT, a transformer trained on galaxy images, as a testbed for studying concept emergence during training, finding that galaxy properties emerge in a fixed difficulty-based sequence, which can inform mechanistic interpretability methods for large language models.

Interpretability research increasingly asks when concepts emerge during training and whether linear probes recover real structure, but in language models these claims are hard to validate because language offers little ground-truth ordering of concepts or relationships among them. We propose the use of astronomical ground truth through AstroPT, a transformer trained on millions of galaxy images, as a calibration testbed. AstroPT is an LLM-like model trained within a domain where the difficulty ordering of concepts and the relations among them are known in advance. Probing frozen representations across checkpoints, layers, model sizes, and objective choices, we find that galaxy properties emerge in a fixed order that tracks their known difficulty---quantities written almost directly into the pixels (band magnitude) become decodable early in training and shallow in the network, while multiband/spectra based and inferred quantities (such as redshift and specific star formation rate) emerge later and deeper. This order is invariant to our tested training objectives, and scales in magnitude but not in sequence with capacity. Our linear probe directions further recover the known physical structure among galaxy properties. Our findings suggest that astronomy offers a controlled sandbox for calibrating mechanistic interpretability methods we otherwise apply to LLMs blind.
Original Article
View Cached Full Text

Cached at: 08/26/26, 03:18 AM

Paper page - What AstroPT knows about galaxies, and what that can teach us about LLMs

Source: https://huggingface.co/papers/2608.22614

Abstract

Using a galaxy-image transformer with known physical concept ordering, the study shows that linear probes recover real structure and that concepts emerge in a fixed difficulty-based sequence during training.

Interpretability research increasingly asks when concepts emerge during training and whetherlinear probesrecover real structure, but in language models these claims are hard to validate because language offers little ground-truth ordering of concepts or relationships among them. We propose the use of astronomical ground truth throughAstroPT, a transformer trained on millions of galaxy images, as a calibration testbed.AstroPTis an LLM-like model trained within a domain where the difficulty ordering of concepts and the relations among them are known in advance. Probingfrozen representationsacrosscheckpoints, layers, model sizes, and objective choices, we find thatgalaxy propertiesemerge in a fixed order that tracks their known difficulty---quantities written almost directly into the pixels (band magnitude) become decodable early in training and shallow in the network, while multiband/spectra based and inferred quantities (such asredshiftandspecific star formation rate) emerge later and deeper. This order is invariant to our tested training objectives, and scales in magnitude but not in sequence with capacity. Our linear probe directions further recover the known physical structure amonggalaxy properties. Our findings suggest that astronomy offers a controlled sandbox for calibratingmechanistic interpretabilitymethods we otherwise apply to LLMs blind.

View arXiv pageView PDFAdd to collection

Get this paper in your agent:

hf papers read 2608\.22614

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2608.22614 in a model README.md to link it from this page.

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2608.22614 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2608.22614 in a Space README.md to link it from this page.

Collections including this paper0

No Collection including this paper

Add this paper to acollectionto link it from this page.

Similar Articles

Understanding Large Language Models

arXiv cs.CL

This chapter reviews current understanding of Large Language Models, discussing their Transformer architecture, emergent capabilities resembling human cognition, and debates about whether LLMs genuinely understand or merely simulate understanding.