JTPRO Framework Improves Tool-Calling Reliability for LLM Agents

Researchers have introduced Joint Tool-Prompt Reflective Optimization (JTPRO), a framework designed to enhance the reliability of large language model (LLM) agents when using large, domain-specific tool inventories. The framework addresses two root causes of tool mis-selection and incorrect parameter instantiation: generic prompts that ignore tool-specific nuances, and underspecified tool schemas lacking guidance on usage and formatting. JTPRO iteratively uses rollout-driven reflection to co-optimize global instructions and per-tool descriptions, improving tool selection and argument instantiation in trace-supervised settings. The approach preserves only tool-local cues needed for accurate performance. The work is detailed in a paper on arXiv (ID: 2604.19821).