Authors: Immanuel Abdi and Akshat Gupta and Micah Mok and Alexander Lu and Nicholas Lee and Gopala Anumanchipalli
Cite as: \cite{abdi2026evolutionary}
arXiv: https://arxiv.org/abs/2601.20861
Summary in three sentences
This paper studies ES for continually learning LLMs and focuses on forgetting over repeated update steps. The arXiv abstract reports that ES can approach GRPO-like task performance while causing substantial forgetting of prior abilities, with ES updates described as less sparse and much larger in norm than GRPO updates. For this project, it is a warning that decomposed ES must track preservation and drift, not just target reward.
Main contribution
The main contribution is an analysis of catastrophic forgetting induced by ES-style LLM updates, including a proposed explanation based on update sparsity and norm.
Key equations and definitions
The key quantities for our work are forgetting curves, update sparsity, and update norm. These should become evaluation diagnostics if the project moves beyond smoke tests.
Empirical results
Every number here must name the section or figure it comes from.
If you cannot locate it in the paper, do not write it.
The arXiv abstract reports math and reasoning evaluations with forgetting curves. No exact performance or forgetting numbers are recorded here until the paper sections are read directly.
Relation to our project
- what we reuse: forgetting and update norm as failure-mode measurements.
- what we dispute: adapter-space or decomposed ES may reduce the dense full-model drift described here, but that needs evidence.
- what we must reproduce: a preservation check on any LLM task, even if the primary reward improves.
Criticism and limitations
The abstract frames ES forgetting as a serious risk, but the details must be read before transferring the conclusion to adapter-only ES. This paper should keep the project honest about negative results.