Tag: In-Context Learning

Sep 21
Zero-Shot Tool Generalization: Evaluating Schema-Only Execution Without Prior In-Context Demos

In early autonomous agent deployments, developers relied heavily on extensive few-shot prompt engineering. To ensure an agent invoked an application programming interface correctly, system prompts were packed with hand-crafted input-output examples, few-shot demonstration trajectories, and explicit step-by-step formatting templates. By embedding past successful runs directly into the context window, developers guided the language model through […]