Machine Learning

Scores Are Not Decisions: Cost-Aware Stopping for Tool Acquisition in LLM Agents

Yicheng Feng, Yan Zhang, Yan Cheng, Wei Qi

Featured August 1, 2026

AI-generated analysis — This is SciGrove's AI interpretation of the paper, not peer-reviewed content. Always refer to the original paper.

Simply

LLM agents often need tools, but picking too many wastes money and resources, while too few means failure. This paper introduces CAM-DF, a smart system that learns exactly when to stop picking tools, balancing usefulness and cost.

In depth

The paper tackles the challenge of cost-aware tool acquisition for LLM agents, where existing tool rankings fail to specify the optimal number of tools to select, especially under heterogeneous costs. The authors introduce CAM-DF, a learned stopping policy that directly optimizes a downstream payoff function, balancing task sufficiency with acquisition costs. CAM-DF trains on the 'decision gap' between stopping now and the best future continuation, weighting errors by their potential payoff loss, which ensures Bayes-alignment with the optimal stopping target.

Key Takeaways

1
The paper formulates tool acquisition as a cost-aware stopping problem over ranked prefixes, enabling optimization for downstream utility while accounting for task sufficiency and heterogeneous costs.
2
It develops a decision-focused objective with Bayes-alignment guarantees, training on the payoff gap between stopping and continuing, and weighting errors by the magnitude of potential payoff loss.
3
The proposed CAM-DF family of policies (including a lightweight variant) provides a deployable, modular control layer that converts existing tool rankings into cost-optimal acquisition decisions without modifying the underlying LLM.

Conceptual Flow

HIGH LEVEL

Methodology: Smart Tool Picking

The system takes a list of tools and their costs, then learns to pick just enough tools to do the job well without wasting resources.

Tool List

Tool Costs

Learn Best Stop Point

Selected Tools

Results: Efficient Tool Use

The new method significantly reduces the number of tools an agent needs to look at, saving costs while keeping tasks successful.

Many Tools Used (Old Way)

Get Same Result With

Fewer Tools Used (New Way)

This breakdown was generated by SciGrove. Get the same analysis — intuition, storyboard, peer review, a runnable prototype and a glossary — on any paper you upload or paste a DOI for.