Self-Improvement of Context, Harness, and Model Weights through Reflective Optimization

Large language models are increasingly adapted to downstream tasks via reinforcement learning methods like GRPO, which often require thousands of rollouts to learn new tasks. We argue that language provides a much richer learning medium: an LLM can reflect on full trajectories (including reasoning, tool calls and errors) to diagnose failures and propose targeted improvements. We introduce GEPA, a reflective prompt optimizer that incorporates this principle outperforming GRPO by up to 20% while using up to 35x fewer rollouts across tasks spanning 5+ domains and also works with black-box models.

Building on this, we then introduce optimize_anything, a unified API that generalizes reflective optimization to arbitrary text parameters. This single system achieves state-of-the-art results across eight fundamentally different areas, including nearly tripling ARC-AGI accuracy via agent architecture discovery, generating CUDA kernels that beat PyTorch and cutting cloud scheduling costs by 40% through policy discovery, establishing LLM-based reflective search as a general-purpose problem-solving paradigm.

Finally, I present Fast-Slow Training (FST), which brings reflective optimization into LLM post-training. FST jointly optimizes model parameters ("slow weights") via RL and textual contexts ("fast weights") via GEPA. Because the fast channel quickly absorbs task-specific nuances, the slow parametric updates are freed to consolidate general reasoning rather than memorizing task details. This yields up to 3x better sample efficiency, a higher performance asymptote with a significantly lower drift from the base model. This reduced drift preserves plasticity for continual learning, allowing FST to adapt sequentially where parameter-only RL stalls.

Broadly, our work advocates a fundamental shift in AI adaptation: replacing task-specific algorithms with diagnostic evaluation, and evolving from parameter-only post-training to the joint optimization of prompts, agent architectures, and model weights.