Stanford TRACE System Trains LLM Agents on Failures

Stanford researchers introduced TRACE, an open-source system that turns recurrent agent failures into targeted training environments. TRACE diagnoses missing capabilities, builds synthetic environments per gap, trains LoRA adapters with GRPO, and composes them via Mixture-of-Experts. On Qwen3-30B-A3B, TRACE improved τ²-Bench by +15.3 points and SWE-bench Verified by +15 points Pass@1.