Create and explore imaginative 3D worlds with our new product Dreamfield (opens in a new tab)

Resolve decoder-logit probe differences among spectral representations

Determine whether continuous spectra, hard category codes, and category-posterior inputs produce different balanced accuracies for decoder-logit probes after frozen decoding.

Background

SPECTRA-VLA evaluates whether spectral information remains accessible after processing by a frozen vision-language-action decoder. Decoder-logit probes are external four-state classifiers trained on the decoder’s first-position action-symbol logits.

Although the paper reports differences in category recognition, within-category spectral recovery, and decoder-KL disturbance under some conditions, it does not resolve whether the three spectral representations yield different decoder-logit probe performance. This leaves the comparative accessibility of their information at the frozen decoder interface unresolved.

References

Decoder-logit probe differences remain unresolved.

— SPECTRA-VLA: Base-Fine Hyperspectral Token Transfer at Frozen Vision-Language-Action Interfaces  (2610.19651 - Hsu, 14 Oct 2026) in Section 5, subsection “Calibrating Recognition and Disturbance”