arXiv 2601.10657

PACEvolve: Enabling Long-Horizon Progress-Aware Consistent Evolution

By Minghao Yan, Bo Peng, et al.

Published 2026-01-15

Citation lineage

Review the prior work and downstream research connected to this paper.

Large Language Models (LLMs) have emerged as powerful operators for evolutionary search, yet the design of efficient search scaffolds remains ad hoc. While promising, current LLM-in-the-loop systems lack a systematic approach to managing the evolutionary process. We identify three distinct failure modes: Context Pollution, where experiment history biases future candidate generation; Mode Collapse, where agents stagn…

View the original paper on arXiv