{
  "id": 4607,
  "url": "https://arxiv.org/abs/2605.09777v2",
  "title": "EvoPref: Multi-Objective Evolutionary Optimization Discovers Diverse LLM Alignments Beyond Gradient Descent",
  "summary": "Gradient-based preference optimization methods for large language model (LLM) alignment suffer from preference collapse, converging to narrow behavioral modes while neglecting preference diversity. We introduce EvoPref, a multi-objective evolutionary algorithm that maintains populations of Low-Rank Adaptation (LoRA) adapters optimized across helpfulness, harmlessness, and honesty objectives using Non-dominated Sorting Genetic Algorithm II (NSGA-II) selection with archive-based diversity preserva",
  "authors": "Dongxin Guo, Jikun Wu, Siu Ming Yiu",
  "category": "research",
  "topics": "safety-alignment",
  "orgs": null,
  "regions": null,
  "published_at": "2026-05-10T21:50:04.000Z",
  "fetched_at": "2026-07-14T16:31:08.355Z",
  "source_slug": "arxiv-ethics",
  "source_name": "arXiv",
  "source_homepage": "https://arxiv.org",
  "ethics_ai_record_url": "https://ethics.ai/record/4607",
  "original_url": "https://arxiv.org/abs/2605.09777v2",
  "evidence_status": "source-only",
  "attribution": "via ethics.ai"
}