Agentic Reasoning and Tool Integration for LLMs via Reinforcement Learning

Large language models (LLMs) have achieved remarkable progress in complexreasoning tasks, yet they remain fundamentally limited by their reliance onstatic internal knowledge and text-only reasoning. Real-world problem solvingoften demands dynamic, multi-step reasoning, adaptive decision making, and theability to interact with external tools and environments. In this work, weintroduce ARTIST (Agentic Reasoning and Tool Integration in Self-improvingTransformers), a unified framework that tightly couples agentic reasoning,reinforcement learning, and tool integration for LLMs. ARTIST enables models toautonomously decide when, how, and which tools to invoke within multi-turnreasoning chains, leveraging outcome-based RL to learn robust strategies fortool use and environment interaction without requiring step-level supervision.Extensive experiments on mathematical reasoning and multi-turn function callingbenchmarks show that ARTIST consistently outperforms state-of-the-artbaselines, with up to 22% absolute improvement over base models and stronggains on the most challenging tasks. Detailed studies and metric analysesreveal that agentic RL training leads to deeper reasoning, more effective tooluse, and higher-quality solutions. Our results establish agentic RL with toolintegration as a powerful new frontier for robust, interpretable, andgeneralizable problem-solving in LLMs.