Progra: Progress-Aware Reinforcement Learning for Multi-Turn Function Calling

Real-world tasks with Large Language Models (LLMs) require multi-turn, multi-step conversations, often involving complex function calls and multiple user interactions. Existing methods either decompose multi-turn trajectories into independent single samples, neglecting task-level structure, or rely on end-to-end reinforcement learning (RL) without explicit progress modeling. To overcome these limitations, we propose PROGRA, a framework that explicitly incorporates progress awareness into LLM training for multi-turn function calling. PROGRA combines a Progress Awareness Generation pipeline to construct training data linking conversation summaries with future plans, and Progress Awareness-Guided RL to condition decisions on this awareness, reducing redundancy and aligning local actions with global task completion. Experiments on two public benchmarks show that PROGRA substantially outperforms prior methods for robust, efficient multi-turn function calling. Our code is available at https://github. com/FatCatCHC/Progra.