Definition
Plain language
A reward that pays an AI more when its tool use matches the kind of task at hand.
As stated in the literature
In ToolCUA, a binary reward signaling whether the agent's use (or non-use) of structured tools matched the task's tool-beneficial label, decoupled from raw success to encourage selective tool invocation.
Also called: tool appropriateness
Why it matters: Without it, agents trained for tool use over-invoke tools on tasks where plain reasoning would be faster and more reliable.
For example, an agent gets credit for opening a calculator tool on a math task and for not opening it when asked the date.
Heard on the show
“The first term is what they call tool appropriateness.”Episode 066 — Why Giving an AI Agent More Tools Can Make It Worse at Using a Computer