Researchers have compared four different methods for adapting AI models to new tasks: zero-shot learning, in-context attention, test-time gradient adaptation, and emitting specialist weights from a hypernetwork. They found that emitting specialist weights is often the most cost-effective option when quality is matched, but it may not be as effective on high-dimensional sequence modeling tasks. The study also showed that emitted specialists can be composed in weight space to create new functions. --- Why it matters: This research matters because it provides insights into how AI models can be efficiently adapted to new tasks, which is crucial for real-world applications such as few-shot learning and transfer learning. Understanding the trade-offs between different adaptation mechanisms can help engineers design more efficient and effective AI systems. Source: https://arxiv.org/abs/2608.21386