Cost-Aware Stopping for LLM Tool Acquisition

A new research paper from arXiv introduces CAM-DF (Cost-Aware Marginal Decision-Focused Stopping), a method for determining how many external tools an LLM agent should acquire. The approach addresses the trade-off between under-informing tasks with too few tools and incurring excessive cost, context load, and privacy exposure with too many. Existing ranking methods fail to account for heterogeneous costs. CAM-DF trains directly on the offline gap between stopping and continuing, using its sign to label decisions and magnitude to weight errors by payoff. A compact variant, CAM-DF-lite, is also proposed. The paper proves that score-only rules are suboptimal under cost heterogeneity.