Researchers develop new training framework to improve LLMs' ability to execute complex multi-step API sequences with real-world data and nuanced reward signals

Current state-of-the-art LLMs struggle with multi-step tool orchestration, particularly when parameter values must be correctly extracted from intermediate outputs

New framework uses a large-scale cache of real API responses to generate realistic training data with controllable complexity, replacing inefficient simulated environments

Introduces 'graduated reward' system that recognizes partial correctness in tool sequences rather than using binary pass/fail signals, improving learning signals

Addresses two critical gaps: existing training environments rely on simple single-turn function calls with synthetic data, and lack feedback for incomplete but partially correct executions