Reward Design Critical for Small-Population LLM Evolutionary Fine-Tuning

Sung Cho, Gyubin Han· July 23, 2026 View original

Summary

Research shows that the effectiveness of Evolutionary Strategies (ES) for fine-tuning large language models with small population sizes is heavily influenced by reward design and normalization, not just population size. Disabling z-score advantage normalization allows binary-reward ES with small populations (N=2) to succeed where normalized variants fail, improving performance across various models and tasks.

A new study investigates the conflicting conclusions regarding population size in Evolutionary Strategies (ES) for fine-tuning large language models (LLMs). While some applications suggest N=1 is sufficient, others require N≈30. The research demonstrates that this discrepancy primarily stems from how rewards are designed and normalized, rather than an intrinsic population limit. Specifically, the study found that z-score advantage normalization can hinder small-population ES, causing it to fail. By disabling this normalization, binary-reward ES with a population size as small as N=2 was able to improve LLM performance on tasks like GSM8K and TREC across models ranging from 0.5B to 7B parameters. This suggests that careful consideration of reward granularity and normalization is key to successfully applying memory-efficient ES for LLM fine-tuning.

Why it matters

This research provides critical insights for optimizing LLM fine-tuning processes, potentially enabling more efficient and cost-effective training, especially for scenarios requiring black-box or discrete rewards.

How to implement this in your domain

  1. 1Review current LLM fine-tuning strategies, especially those using ES, for reward design and normalization practices.
  2. 2Experiment with disabling or adjusting reward normalization in ES fine-tuning for specific tasks.
  3. 3Investigate the impact of reward granularity on training stability and performance with small populations.
  4. 4Consider ES as a viable, memory-efficient alternative for fine-tuning when black-box rewards are necessary.

Who benefits

AI/ML DevelopmentSoftware EngineeringResearch & DevelopmentCloud Computing

Key takeaways

  • Reward design and normalization are more critical than population size for small-N ES in LLM fine-tuning.
  • Z-score advantage normalization can be detrimental to small-population ES performance.
  • Disabling normalization can enable effective binary-reward ES with very small populations.
  • This finding can lead to more memory-efficient and cost-effective LLM fine-tuning.

Original post by Sung Cho, Gyubin Han

"arXiv:2607.19408v1 Announce Type: new Abstract: Using Evolutionary Strategies (ES) for fine-tuning large language models is attractive because it is memory-efficient, parallel, and compatible with black-box or discrete rewards. Yet its population-size conclusions conflict sharply…"

View on X

Originally posted by Sung Cho, Gyubin Han on X · view source

Want to go deeper?

Turn these trends into skills with Learnijoy's hands-on AI & tech courses.

Explore courses