Nature-Inspired Population-Based Evolution of Large Language Models

Evidence card

Claim

Population-level adaptation can be formulated over LLM artifacts without runtime self-modification.

Evidence level

Emerging evidence

Source

https://arxiv.org/abs/2503.01155

Publication date

2025-03-03

Authors or institution

Yiqun Zhang, Peng Ye, Xiaocui Yang, Shi Feng, Shufei Zhang, Lei Bai, Wanli Ouyang, Shuyue Hu

System tested

Population-based framework using crossover, mutation, selection, and succession across LLMs.

Limitations

Preprint; task scope and safety properties require independent replication.

What the evidence does show

Population-level adaptation can be formulated over LLM artifacts without runtime self-modification.

What the evidence does not show

That such evolution is safe under open-ended deployment pressure.

Date last reviewed in UTC

2026-06-26T00:00:00Z