​
Abstract

This case study demonstrates the power of layering a hyper-efficient online optimization process (GEPA) on top of a powerful, RL-trained foundation model. By using a gemini/gemini-flash-2.5 reflection agent to evolve the prompts of our ATLAS-8B-Thinking teacher model, we achieved a +165% performance improvement on the student model in approximately 2 hours and at a cost of only ~$10 in API inference.

+165% Performance

Measured by the increase in the Pareto front score, reflecting robust and diverse solutions.

~2 Hours to Optimize

Rapid adaptation without the need for expensive, long-running training jobs.
ATLAS Online Learning Improvement

​
The Hybrid Architecture: Offline RL + Online Optimization

This experiment showcases a hybrid approach that combines the best of offline and online learning:
  1. Offline RL Foundation (ATLAS): The ATLAS-8B-Thinking teacher model is first trained offline using our Reinforced Continual Learning (RCL) process. This builds a deep, generalizable foundation of reasoning and pedagogical skills.
  2. Online Optimization (GEPA): We then use a “reflection agent” to analyze student failures on a live task and propose targeted “reflective mutations” to the teacher’s prompts. This rapidly specializes the teacher for the specific task domain.

​
The Evolution of a Prompt

The optimization process intelligently evolves prompts from a generic instruction to a sophisticated teaching framework.

​
Initial Prompt (Score: -0.2)

The initial prompt was too generic and actually harmed performance: 
You are solving math word problems. Think step by step and
show your work clearly to arrive at the correct answer.

​
Final Evolved Prompt (Score: 1.479)

After reflective mutation, the prompt became a detailed, multi-step teaching strategy: 
You are an expert teacher helping a student solve a math problem.
Your primary goal is to provide focused, adaptive teaching that
directly addresses the student's specific misconceptions or gaps...

1. Analyze the Student's Approach Step-by-Step
2. Identify Specific Issues, Misconceptions, or Gaps
3. Provide Focused and Actionable Teaching...
The reflection agent identified key failure modes, such as the student providing solutions directly instead of reasoning, and evolved the prompt to enforce a diagnostic process.

​
Real-World Impact

This hybrid architecture provides a powerful new paradigm for agent builders. It delivers the reliability of a model thoroughly trained with RL, plus the speed and task-specificity of online learning, without requiring massive compute for every new adaptation.

​
Next Steps

Full Technical Breakdown

Read the original article for more details.

Online Optimization Guide

Learn how to run this process yourself.