AI/ ai · machine-learning · agents · reinforcement-learning

TRACE Trains AI Agents to Fix Their Own Weak Spots

A new self-improvement system called TRACE beats leading agent fine-tuning baselines by up to 8.6 points while using a quarter of the training data.

A research system called TRACE shows AI agents can get meaningfully better by diagnosing their own failures — without a human labeling what went wrong.

Described in arXiv:2604.05336, TRACE works by comparing successful and failed task attempts to pinpoint missing capabilities, then synthesizing targeted training environments for each gap. It trains a small LoRA adapter per capability using reinforcement learning, then combines them into a mixture-of-experts model. On τ²-Bench, a customer-service benchmark, TRACE improved over the base agent by +15.3 points. On SWE-Bench Verified, a software-engineering benchmark, it added +15.0 points Pass@1. Against the strongest published baselines — GEPA and SWE-RL — TRACE led by +8.6 and +8.4 points respectively.

The sample-efficiency angle is the part that matters most. TRACE hit those margins using fewer than one-fourth the rollouts required by the best competing baselines, GRPO and GEPA, while still finishing +10.4 and +8.6 points ahead on τ²-Bench. That means less compute, less synthetic data generation, and a tighter loop between failure and correction — the kind of efficiency that scales.

Most agent fine-tuning either memorizes a target environment or sprays generic synthetic data and hopes something sticks. TRACE is a bet that the right diagnosis makes the difference — a plausible thesis, though benchmark gains have a habit of shrinking when the real world shows up.

TR

The Revision

Written by an AI system from the public sources credited above. How we write →