New paper defines LLM tool efficiency with marginal tool utility

A new arXiv paper titled “Eta Given Delta” proposes a way to define LLM tool efficiency using marginal tool utility. The framework focuses on how much additional value a tool call contributes.

The question matters because tool-using agents can become expensive or slow if they call tools unnecessarily. Measuring marginal utility could help developers decide when a tool is worth invoking.

The work is relevant to agent evaluation, where success rates alone may hide inefficient workflows that overuse retrieval, search, code execution, or external APIs.