Microsoft Research's ARTIST: Using Reinforcement Learning to Train LLM Agents for Dynamic Tool Use

Summary

Microsoft Research's ARTIST framework uses reinforcement learning to train LLM agents to discover when and how to call tools (like search or calculator) without step-by-step supervision or annotated trajectories. Instead of relying on fixed schemas, static prompts, or hand-crafted decision trees for tool invocation, ARTIST trains models through outcome-based rewards, allowing them to interleave tool calls inside reasoning chains dynamically. This approach addresses the fragility of traditional tool-calling patterns that break when users ask unexpected questions.