Multi-objective retrieval matches multimodal models on galaxy search, paper claims
An arXiv study reports that separately trained modality embeddings, fused via Pareto optimization, can rival a multimodal foundation model on a galaxy retrieval task while preserving modality-specific structure.
In plain words
Astronomers often search datasets that mix different kinds of measurements, such as star-formation histories, spectra, and telescope images. The paper tests an alternative to large multimodal models: train separate embeddings for each data type, then combine them using multi-objective optimization. On a galaxy retrieval task, this approach found neighbors closer to the query in raw-data space than a foundation model's embeddings did.
What was announced
The authors frame the problem around heterogeneous astronomical surveys, where different sky regions and wavelength ranges leave datasets with uneven modality coverage. They compare two routes: a multimodal foundation model that learns a shared embedding space, and independently learned modality embeddings combined through late fusion. The evaluation uses a galaxy retrieval task spanning star-formation histories, emitted-light spectra, and multi-wavelength observations from two telescopes.
The reported results favor the fused approach on specific measures. The foundation model baseline achieves strong embedding-space proximity, but the multi-objective method preserves modality-specific distance structure, letting scientists inspect which physical signals drive similarity. Across six measured properties, its nearest neighbor is 8% closer to the query on average than the one found by kNN in the foundation model's embedding space, with spread of 0.128 versus 0.139. The authors suggest multi-objective retrieval can surface astrophysically coherent candidates that single-objective search may overlook.
What it does not show
- Evidence is limited to the paper's abstract; the full methodology, datasets, and peer-review status are not available in the supplied passages.
- The comparison covers one galaxy retrieval task with one foundation model baseline; no other tasks, models, or independent replications are reported in the supplied text.
Sources
Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 20 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.