Intel / video

Beating RL With Reflection: GEPA and Optimize Anything — Lakshya A. Agrawal, GEPA

Source: youtube.com
Author: AI Engineer
Date: Published Sep 26, 2026

From the Daily Brief

Lakshya A. Agrawal, the UC Berkeley PhD student who created GEPA, argues that reinforcement learning throws away most of what a rollout contains by collapsing it into one score. GEPA keeps the text. A model reads the full trace, including chains of thought, tool calls and error messages, and proposes a better prompt, while a Pareto pool of candidates keeps the search from settling on a local optimum. In his talk, Agrawal says one round of reflection on three examples doubled the gains GRPO reached after 25,000 rollouts. The successor, Optimize Anything, applies the same loop to anything that can be written as text and scored. He reports it grew a four-line program into a six-step agent that lifted Gemini Flash on ARC-AGI from 32.5% to 89.5%, raised an AMD NPU coding agent from 4% to 30%, and took a GPT-5 mini coding agent on Go issues from 24% to 93% with learned skills that also transferred to Claude Sonnet. These are the speaker's own numbers, not an independent replication. For teams with scored tasks and readable traces, trying reflective prompt search before a costly RL run looks like the cheaper first experiment.

Read the 2026-09-27 Brief →

Context

GEPA uses the details of failed attempts to propose better prompts and code. Its optimization loop runs a candidate, records information such as error messages and execution logs, and asks a to diagnose the failure and suggest a revision. The project documentation describes retaining several candidates that succeed on different task subsets, allowing the search to build on more than one promising approach.

The talk's central idea extends this process beyond prompts through Optimize Anything. Developers supply something editable as text and a way to evaluate it; the system searches for better versions. The useful mechanism is turning detailed feedback into targeted changes, with the evaluator defining what improvement means.

Key takeaways · AI-distilled

Agrawal reports that one round of reflection on three examples let GEPA double the gains GRPO reached after 25,000 rollouts. RL squeezes a rollout into one score; GEPA has a model read the full trace, including reasoning, tool calls and errors, and write a better prompt.
GEPA keeps a Pareto pool of candidates rather than one running best, which Agrawal says keeps the search out of local optima.
Optimize Anything applies the idea to anything expressible as text with a score. He shows it growing a four-line program into a six-step that lifts Gemini Flash on ARC-AGI from 32.5% to 89.5%, and an AMD NPU coding agent from 4% to 30%.
Learned skills raised a GPT-5 mini coding agent on Go issues from 24% to 93% and carried over to Claude Sonnet. He also cites Databricks tuning an open model to beat Claude Opus at 90x lower cost.

Terms in this piece · Glossary

LLM — A large language model — the neural network behind tools like Claude and ChatGPT, trained on huge amounts of text to predict what comes next.
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.

Read the source www.youtube.com