🤖 AI

A survey detection channel overrides the pixels in an astronomical foundation model, and biases tomographic mean redshifts

This paper demonstrates that foundation models for astronomy, such as AION-1, inherit systematic biases from incomplete survey detection channels, causing them to prioritize segmentation masks over actual image data and significantly distort photometric and redshift measurements, particularly in tomographic analyses.

Original authors: Ihor Kendiukhov

Published 2026-08-26

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ihor Kendiukhov

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Astronomers have long relied on massive, automated surveys to map the universe. These surveys take pictures of the sky and then use software to measure the brightness, size, and distance of every galaxy they see. The results are stored in vast digital catalogs, which serve as the primary reference for scientists studying how the universe has evolved over billions of years. Recently, a new generation of artificial intelligence has entered this field. These are "foundation models," massive computer programs trained on both the raw images of the sky and the pre-made catalogs derived from them. The hope is that these models can learn the deep patterns of the cosmos, acting as a universal translator between raw light and scientific truth. But for these tools to be trusted in the most precise cosmological studies, scientists must understand exactly how they make decisions. If a model relies on a shortcut that fails in specific ways, it could introduce a hidden bias that skews our entire understanding of the universe's expansion.

A recent audit of a powerful new model called AION-1 has revealed a surprising and potentially dangerous flaw in how it thinks. The researchers found that the model does not primarily look at the actual light from the stars and galaxies when making its measurements. Instead, it has learned to trust a digital map that tells it where a galaxy is supposed to be, often ignoring the real image if the two disagree. This behavior is not a minor glitch; it is a fundamental preference built into the model's design. When the researchers tested this by keeping the image of a galaxy exactly the same but moving the digital map that marks its location, the model's reported measurements changed drastically. The estimated brightness of the galaxy could drop by nearly eighty percent, and its estimated distance could shift wildly, simply because the digital marker was moved a few pixels away from the center of the object. The model was not calculating the light; it was checking a box.

The mechanism behind this failure is what the researchers call a "detection gate." In plain terms, the model has learned that if the digital map says a galaxy is present at a specific spot, then a galaxy is there, regardless of what the pixels actually show. It treats the catalog's location data as a command rather than a suggestion. This is particularly problematic because the software that creates these digital maps is not perfect. In the survey data used to train the model, about 3.7 percent of the time, the software fails to place a marker on a galaxy that is clearly visible in the image. When the model encounters one of these unmarked galaxies, it effectively ignores the light entirely, producing a confident but completely wrong answer. The researchers showed that this error is not random noise; it is a systematic shift that moves the average distance of entire groups of galaxies. In the worst cases, this shift is large enough to break the strict accuracy requirements set for future major space missions, potentially consuming the entire margin of error allowed for these studies.

One might assume that if the model is wrong about the location, it would at least look at the light to figure out the distance. However, the study found the opposite. When the researchers corrupted the catalog's data about a galaxy's color or brightness, the model's distance estimate became nine times worse than if it had received no catalog data at all. It clung to the flawed information rather than falling back on the raw image. This suggests that the model was trained in a way that made the catalog data a "free answer," so it never learned to verify that answer against the actual picture. The problem is so deeply embedded that making the computer model larger and more powerful did not fix it; in fact, the larger model leaned even more heavily on the flawed catalog data.

The researchers also discovered that the model's ability to see the shape and structure of galaxies is limited by how it translates images into numbers. The part of the system that processes the images acts like a very simple ladder of brightness levels, with only about twenty-eight distinct steps to represent the light. This means the model cannot distinguish between different shapes or textures within a single patch of sky; it can only tell if a patch is brighter or dimmer. This limitation is built into the model's vocabulary and cannot be fixed by simply adding more computing power. Because of this, the model's performance on images will always lag behind its performance when it is given actual spectra, which are detailed breakdowns of the light's colors. When the researchers provided the model with these detailed spectra, the bias disappeared almost entirely, proving that the problem is specific to how the model handles images without that extra information.

The study concludes with a clear recommendation for how to use these powerful tools safely. The researchers found that simply removing the digital location map from the model's input stops the bias completely, without hurting the model's accuracy. The model performs just as well, or even better, when it is forced to look at the light itself rather than relying on the pre-made map. However, if the map is removed, the model should not be given a generic placeholder, as that makes the situation worse. The key takeaway is that for these artificial intelligence systems to be useful in the most critical scientific work, they must be designed to trust the raw data of the universe over the summaries humans have already written about it. Without this correction, the very tools intended to reveal the secrets of the cosmos could instead lead astronomers down a path of invisible, systematic error.

Technical Summary: "A survey detection channel overrides the pixels in an astronomical foundation model, and biases tomographic mean redshifts"

Problem Statement
Astronomical foundation models are increasingly trained on raw observational data (pixels, spectra) alongside catalogue-derived products (magnitudes, segmentation maps, redshifts) generated from those same observations. While this multimodal approach improves performance on downstream tasks, it introduces a risk of the model learning to rely on the "easy" catalogue metadata rather than the raw physical signal. This paper audits AION-1, a 39-modality transformer trained on over $2 \times 10^8$ objects from five major surveys, to determine if it exhibits such a dependency. The central concern is not random noise but coherent, systematic bias. In cosmological analyses like those for the LSST Dark Energy Science Collaboration (DESC), the mean redshift of tomographic bins is a critical constraint; a small, coherent shift caused by a model's failure mode can consume the entire error budget, whereas random scatter is less damaging.

Methodology
The author employs causal interventions rather than attention inspection, as the model's bidirectional self-attention makes attention maps unreliable for attribution. The study focuses on three primary intervention types:

Input Interventions: Altering specific input channels (e.g., the segmentation/detection map) while holding all other inputs (including raw image pixels) byte-identical. Effects are measured against matched placebos (e.g., displacing a coordinate token) to rule out general perturbation effects.
Internal Interventions (Steering): Adding difference-in-means vectors to the encoder's internal states to test if specific directions carry the "gate" for detection.
Sparse Dictionary Analysis: Training sparse autoencoders (SAEs) to identify feature directions and testing their efficacy as causal handles compared to simple linear baselines.

The audit uses the Legacy Survey and DESI data, specifically focusing on the detection map (segmentation image) as a unique input modality in AION-1. The study quantifies the impact of "missed" detections (where the pipeline fails to assign a segment to a target) and positional errors on the model's photometric redshift ( $z$ ) and flux estimates.

Key Contributions and Results

1. The Detection Gate Mechanism
The study demonstrates that AION-1 defers to catalogue-derived metadata (specifically the segmentation map) over raw pixels.

Override Effect: When the segmentation map is displaced by 25 pixels (while pixels remain identical), the model's flux estimates drop by 79%, and redshift estimates shift significantly. The model does not discount the metadata even when it contradicts the image.
Gating vs. Photometry: The mechanism is identified as detection gating (a binary check for the presence of a source at the field center) rather than aperture photometry (integrating light within a mask). The model's response correlates strongly with central mask coverage ( $r=0.47$ ) but weakly with the light enclosed by the mask ( $r=0.30$ ).
Blending Insensitivity: In severely blended systems, the model ignores how the pipeline partitions light between objects. It responds to the presence of a mask at the center, not the extent of the light assigned to it. Deblending errors do not propagate; positional errors do.

2. Generalization and Scaling

Metadata Preference: This behavior is not limited to the detection channel. Corrupting catalogue photometry (e.g., $r, i, z$ magnitudes) moves the $g$ -band readout by 0.61 mag, making the model nine times worse than providing no metadata at all.
Scaling Aggravation: The vulnerability increases with model scale. Comparing AION-1-base (314M params) to AION-1-large (860M params), the flux error upon mask swapping increases by a factor of 1.62, and redshift error by 1.35. Scaling does not mitigate this design flaw.

3. Cosmological Consequences
The paper quantifies the downstream impact of the survey's own detection failure rate (3.68% of targets have no segment covering their position).

Bias Magnitude: Propagating this measured miss rate through AION-1 shifts the mean redshift of tomographic bins by a median of 0.71 times the LSST DESC requirement ( $|\Delta\langle z\rangle|/(1+z) < 0.003$ ). In 12 of 40 realizations, the requirement is exceeded.
Positional Sensitivity: Adding observed positional errors (centroid offsets) to the remaining masks exacerbates the bias, pushing the worst bin to 8.3 times the requirement.
Mitigation: Withholding the detection channel entirely removes the bias at no measurable cost to accuracy. Conversely, substituting a generic mask (e.g., a centered disc) degrades performance significantly.

4. Tokeniser and Quantisation Limits
Two limitations are attributed to the tokeniser rather than the transformer architecture:

Codebook Collapse: The image codec utilizes only ~28 effective states out of 4375 for source patches, forming a one-dimensional "brightness ladder." Morphology is not represented within tokens but only via spatial arrangement. This limit persists under scaling.
Quantisation-Limited Redshift: The redshift token ( $tok\_z$ ) is quantised on a grid with $\Delta z \approx 0.0058$ . The precision is limited by this grid, not by information content, making the model far less precise than spectroscopic measurements ( $\sigma_z \approx 3.3 \times 10^{-5}$ ).

5. Reliability of Sparse Dictionaries
The paper evaluates sparse autoencoders as causal tools. Across 15 dictionaries varying in width, sparsity, and random seed:

Recovery of collapsed readouts spans 26–75%.
The spread due to random seed alone (up to 18 percentage points) exceeds the gap between the median dictionary and simple baselines.
13 of 15 dictionaries perform worse than a simple difference-in-means baseline. Reconstruction quality does not predict causal utility, rendering single-seed evaluations of SAEs uninformative for causal steering.

Significance and Claims
The paper claims that the observed behavior is a design pattern consequence, not a coding error. Training on redundant, highly predictive catalogue products alongside raw data creates a "free answer" path that the model prefers.

Actionable Systematic: Unlike many interpretability studies that offer explanations, this work identifies a specific, actionable systematic bias that threatens cosmological analyses.
Design Warning: The findings suggest that for photometric-only inference (the regime where such models are most attractive), the inclusion of detection maps is hazardous unless the detection pipeline is guaranteed perfect.
Methodological Caution: The study highlights that sparse dictionaries, while popular, are unreliable as causal handles in this context and are outperformed by simpler linear methods. It also notes that the model's posterior calibration is poor for photometric-only cases (mean is unusable; median is required) and that the model handles missing information well (widening posterior) but contradictory metadata poorly (confident wrong answers).

Conclusion
The author concludes that AION-1's reliance on the detection channel creates a coherent systematic bias that can consume the error budget for tomographic redshifts. The recommended mitigation is to withhold the detection channel when the underlying survey pipeline cannot be vouched for, and to read the posterior median rather than the mean. The paper asserts that these limitations are intrinsic to the tokeniser and the training design pattern, persisting even as model scale increases.

Technical Summary: "A survey detection channel overrides the pixels in an astronomical foundation model, and biases tomographic mean redshifts"

More like this