Cut 90% of Fine-Tuning Cost—Still Beat Baselines on Text and Vision Benchmarks

Multimodal hyper-embeddings enable strong cross-task transfer; tiny tuned modules match or beat full fine-tunes on unseen domains.