隐性智能——评估代理在用户未言明内容上的表现

Real-world requests to AI agents are fundamentally underspecified, relying on shared context and unstated constraints that speakers expect listeners to infer. Current benchmarks for agents focus on explicit instruction-following but fail to evaluate their ability to reason about implicit requirements, including accessibility needs, privacy boundaries, catastrophic risks, and contextual constraints. This research presents the Implicit Intelligence evaluation framework, which tests whether AI agents can move beyond mere prompt-following to become genuine goal-fulfillers. Paired with Agent-as-a-World (AaW), this framework allows interactive worlds to be defined in human-readable YAML files and simulated by language models. The scenarios highlight the agents' capabilities in addressing implicit demands.