Act Wisely: Meta-Cognitive Training Teaches Multimodal Agents When NOT to Use Tools — Reducing Unnecessary Tool Calls

arXiv·medium signal

Yan et al. present a meta-cognitive training approach for agentic multimodal models that addresses a critical production problem: current agents over-use tools, making unnecessary API calls that increase latency and cost. The 'Act Wisely' framework trains agents to evaluate whether a tool call is actually needed before executing it, rather than defaulting to tool use for every query. For builders running agent workloads at scale, unnecessary tool calls directly impact cost and reliability — this is the first systematic approach to training tool-use restraint.