May 18, 2026 Learning Fast and Slow: A New Recipe for Adapting LLMs Without Forgetting May 18, 2026 Meta-Harness: End-to-End Optimization of Model Harnesses