Scores Are Not Decisions: Cost-Aware Stopping for Tool Acquisition in LLM Agents
CAM-DF decides when an agent should stop adding tools, not just which tools rank highest.
The paper frames tool acquisition as a cost-aware stopping problem over ranked tool lists. Its training target uses the offline gap between stopping now and continuing, weighting mistakes by the payoff at stake. Across 1,343 tasks in five domains, CAM-DF performs best under heterogeneous costs and high cost pressure, especially when rankings are weaker. In live execution, it exposes agents to 37% fewer tools than full access while keeping task success comparable. Source: ArXiv · AI/CL/LG's note.
The paper frames tool acquisition as a cost-aware stopping problem over ranked tool lists. Its training target uses the offline gap between stopping now and continuing, weighting mistakes by the payoff at stake. Across 1,343 tasks in five domains, CAM-DF performs best under heterogeneous costs and high cost pressure, especially when rankings are weaker. In live execution, it exposes agents to 37% fewer tools than full access while keeping task success comparable. Source: ArXiv · AI/CL/LG's note.
score 4