We propose Test-Time Spectrum-Aware Latent Steering (STS), a lightweight adaptation that learns a few per-sample steering parameters in latent space to align visual and text embeddings and boost zero-shot robustness.