OmniSpectra: a wavelength-native spectral foundation model that skips destructive resampling
Md Khairul Islam · hf · 2026-10-29
A University of Virginia and SLAC team released OmniSpectra, a spectral foundation model (arXiv, Oct 2026) that uses continuous coordinates and physics-grounded decoders instead of integer pixel indexing. It maps the multiplicative scaling from cosmological redshift into a uniform additive shift under continuous rotary position embeddings, so attention geometry no longer depends on the sampling grid.
This addresses the structural mismatch where instruments sample the same wavelengths at different pixel densities, enabling cross-instrument representation learning without destructive resampling — with high performance from far fewer parameters.
More from Research
- AI Now Discovers Faster Than Institutions Can Verify: The 200M-Protein Integrity Paradox — mikeflache · 2026-10-29
- FrontierCS 2 reshuffles rankings: Claude Opus 5.5 tops FCI as DeepSeek 1017 degrades — teortaxesTex · 2026-10-29
- Constitutional grafting: character training for open models under $500, skipping $440k of compute — voooooogel · 2026-10-29
- Transformer layers turned into a full-fledged symbolic CPU — yshan2u · 2026-10-29
- NanoDreamer: A 750K-parameter world-model agent trained on a MacBook M5 Pro plays 5 MinAtar games — CatAstro_Piyush · 2026-10-29
- Why You Can't Just Remove Positional Embeddings: Self-Attention Is Permutation Equivariant — cneuralnetwork · 2026-10-29