Stanford Researchers Introduce TRACE: A Capability-Targeted Agentic Training System That Turns Recurrent Agent Failures Into Synthetic RL Environment

Short summary

Stanford researchers introduce TRACE, a system that diagnoses recurring agent failures from trajectory data, synthesizes per-capability RL training environments, trains LoRA adapters, and routes tokens across experts. It improves τ²-Bench by +15.3 points and reaches 73.2% Pass@1 on SWE-bench Verified. The article is extremely brief and essentially restates the abstract.

• •TRACE turns recurring agent failures into synthetic RL training environments
• •Trains one LoRA adapter per capability and routes tokens across experts
• •+15.3 points on τ²-Bench and 73.2% Pass@1 on SWE-bench Verified

Generated with AI, which can make mistakes.