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

Mechanistic interpretability research seeks to understand when concepts emerge during model training and whether linear probes accurately reflect true underlying data structure. However, in large language models (LLMs), these claims are notoriously difficult to validate because natural language lacks a strict, known ground-truth hierarchy of concepts and relationships. To overcome this limitation, the authors propose using astronomy as a controlled calibration sandbox via AstroPT, a transformer-based foundation model trained on millions of galaxy images.

AstroPT mimics the architecture and training objectives of modern LLMs—such as causal autoregressive (GPT-like) and masked autoencoding (BERT-like) approaches—by treating image patches as tokens. Because astronomy possesses well-established physical laws, empirical scaling relations, and a clear hierarchy of conceptual difficulty, it provides a rigorous ground-truth testbed. Simple pixel-level properties like apparent magnitude are straightforward, whereas integrated or inferred properties like redshift and specific star formation rate (sSFR) require complex physical context and multi-band data. This setup allows researchers to test whether hidden representations actually internalize known scientific geometry.

By probing frozen checkpoints across various network layers, model sizes, and training objectives, the study tracks the decodability of three key galaxy properties: r-band magnitude, redshift, and sSFR. The results reveal that these properties do not emerge simultaneously. Instead, they follow a strict difficulty-ordered sequence. Quantities tied directly to raw pixel values become linearly decodable early in training and in the shallowest layers of the network. Conversely, integrated quantities like redshift emerge later in training and deeper in the architecture, while highly indirect properties like sSFR remain weak within a single epoch.

Crucially, this emergence sequence remains invariant to changes in the training objective (comparing autoregressive and masked autoencoding objectives). Furthermore, scaling model capacity from 1 million to 100 million parameters increases the magnitude of probe performance (R-squared scores) but does not alter the sequence of property emergence. This demonstrates that the ordering is a fundamental property of learning the data structure rather than an artifact of a specific training recipe.