Evolutionary Strategies lead to Catastrophic Forgetting in LLMs

arxiv:2601.20861

Original source · Paper JSON

Post-training · 2026-01-28

Cites (4) · Cited by (0)

Complete cleaned Markdown · Source Markdown including bibliography

Read in parts

1. Part 1
2. Part 2

Contents

• Evolutionary Strategies lead to Catastrophic Forgetting in LLMs
• Abstract
• 1 Introduction
• 2 Related Work
• 3 Experiments
• 3.1 ES vs GRPO Comparison
• 3.2 ES and Catastrophic Forgetting
• 3.3 Dissecting ES Updates: Norm and Sparsity
• Norm.
• Sparsity.
• 4 Conclusion
• Limitations
• Appendix A Appendix
• A.1 Algorithmic Overview and Analogy Between ES and GRPO
• A.1.1 ES Algorithm Overview
• A.1.2 ES Algorithm Overview
• A.1.3 Analogy Between ES Population Size and GRPO Rollout Count
• A.2 Implementation Details
• A.2.1 GRPO
• A.2.2 ES
• A.2.3 Reward functions
• A.3 Hyperparameter Values
• A.4 Additional Experiments
• A.4.1 Catastrophic Forgetting and KL
• Instructions for reporting errors