Justy: Stanford just figured out how to turn agent failures into a training system that actually fixes the right thing.
Justy: Most RL training for agents is broad. You throw rollouts at a model and hope it generalizes. TRACE does something narrower… it diagnoses which specific capabilities are missing.
Justy: It splits rollouts into pass and fail, then finds the gaps that consistently separate them. Each gap gets its own synthetic environment — auto-verifiable, seeded, one per weakness.
Justy: Then a small LoRA adapter trains on that environment alone. Base model stays frozen. You're only updating the delta for that capability.
Justy: At inference, a learned gate routes each token — top-one — to whichever expert covers that capability. So the model isn't generalizing across everything… it's pulling the right specialist per token.
Justy: On τ²-Bench with Qwen3-30B, this beats both prompt optimization and single-adapter baselines. That's a meaningful gap.
Justy: The paper's on arXiv — 2604.05336 — and the code is linked in the Marktechpost writeup. Worth reading the capability analysis section specifically… the threshold math for which gaps get retained is the sharpest part.
Justy: Failure as a training signal, routed at the token level — that's where agentic reliability is heading.