Stanford researchers introduce TRACE system for agentic training

TRACE (Turning Recurrent Agent failures into Capability-targeted training Environments) diagnoses missing capabilities in agentic LLMs and trains on synthetic RL environments built from recurring failures. The system aims to address repeated failures by targeting specific capability gaps.