Daily Digest — 2026-09-27

Lakshya A. Agrawal introduces Reflective Optimization (GEPA), addressing the challenge of teaching AI new tasks, particularly with limited data and expensive rollouts, a common bottleneck in current AI development. Traditional methods like pretraining, supervised fine-tuning (SFT), and reinforcement learning (RL) are highly data-intensive, requiring trillions of tokens, thousands of labeled examples, or hundreds of thousands of rollouts, respectively.

GEPA proposes a shift from numerical reward signals and weight updates to text-space optimization. The core idea is twofold:

1. Reflect on full traces: Instead of a simple 0/1 reward, GEPA allows an LLM to examine the complete trace of a rollout, including intermediate outputs, tool invocations, environment responses (e.g., error messages), and chains of thought. This provides rich, diagnostic textual feedback. 2. Update prompts as natural language: Rather than small, incremental weight updates via gradient descent, GEPA updates prompts. A single natural language change in a prompt can lead to a substantial change in the agent's behavior, offering high sample efficiency compared to traditional methods.

The GEPA algorithm operates as a loop: 1. Run: Execute the current system (prompt/agent) on a few examples and collect feedback, including intermediate outputs and diagnostic information. 2. Reflect: An LLM or agent reads this collected feedback and proposes a better prompt. 3. Keep a Pareto Pool: Instead of only retaining the single 'best' candidate (which risks getting stuck in local optima), GEPA maintains a Pareto pool of diverse candidates. This strategy allows for broader exploration of the solution space and avoids premature convergence to suboptimal solutions. The speaker demonstrates that this Pareto-based candidate sampling leads to significantly better and faster improvements (e.g., twice the performance gains of GRPO with minimal rollouts in multi-hop QA tasks).

GEPA has shown impressive results across various benchmarks and applications:

Multi-Hop Question Answering (QA): GEPA achieved twice the performance gains of GRPO (a leading RL technique) with just three data points and one round of reflection on HotpotQA and HoVer using a Qwen3 8B model. Continuous running further amplified this gap. System Specification Generation: GEPA can generate complete, detailed system specifications for tasks like multi-hop QA. For example, it produced a prompt for the second hop of a QA system that outlined key observations, lessons, and practical strategies for query refinement and output formatting, which human engineering teams typically spend weeks to discover. Mathematical Reasoning: When applied to the LiveBench Math dataset, GEPA optimized GPT-4.1 mini to outperform GPT-4.1 (a larger, more capable model) by distilling complex reasoning and formatting rules into the prompt. The optimized prompt significantly enhanced accuracy from ~58% to ~64%. Hardware Accelerator Programming: For programming a new AMD Accelerator (NPU XDNN A2) with a completely new API and almost no online documentation, GEPA took an existing agent's performance from 4.25% to 30.52% (a 7x improvement) by optimizing the prompt. This highlights GEPA's ability to discover effective solutions in domains with scarce knowledge resources. Agent Architecture Discovery (ARC-AGI): GEPA was used to evolve the architecture of an agent to solve ARC-AGI problems. Starting with a simple 4-line Python program (a chain-of-thought LLM), GEPA autonomously discovered a sophisticated 6-step agent that performs rule hypothesis induction, code synthesis, execution tracing, and debugging. This improved Gemini Flash's ARC-AGI accuracy from 32.5% to 89.5% in just 16 rounds of reflection, without human intervention. Learning Agent Skills (Claude Code): GEPA optimized the `SKILL.md` files used by coding agents like Mini-SWE-Agent (GPT-5 mini) on two Go repositories. It boosted issue resolution rates from 39% to 83% (a ~3x jump) and from 24% to 99% respectively. Crucially, these skills, optimized cheaply on a smaller GPT-5 mini agent, could then be transferred to a larger model like Claude Sonnet 4.5, resulting in 100% issue resolution with a 50% reduction in execution time (and thus token consumption). * Complex Information Extraction (Databricks): Databricks utilized GEPA to optimize enterprise agents for complex information extraction, achieving a 90x cost reduction while outperforming Claude Opus. This demonstrates GEPA's effectiveness even when optimizing models that are already highly capable.

The speaker emphasizes the `optimize_anything` API's universality: "If you can write it as text and score it, GEPA can optimize it." This applies to prompts, agent harnesses (entire codebases), kernels, numeric parameters, algorithms, and more. The library is designed to be easy to use, framework-agnostic (plugs into DSPy, LangChain, Pydantic, etc.), model-agnostic (works with OpenAI, Claude, Gemini, local models), and has zero hard dependencies. The paradigm enables continual learning by fostering small model drifts at higher accuracy, similar to how humans continuously learn and adapt.

GEPA has already seen adoption in production by companies like AWS, Uber, Dropbox, Databricks, Open AI, and Meta, with positive feedback from industry leaders. The project is fully open-source and available on GitHub, encouraging widespread adoption and further development in optimizing text-based parameters across diverse AI systems.