{
  "id": 3315,
  "url": "https://arxiv.org/abs/2606.01561v2",
  "title": "S-SPPO: Semantic-Calibrated Self-Play Preference Optimization",
  "summary": "Aligning Large Language Models (LLMs) with human preferences is often formulated via Direct Preference Optimization (DPO). However, the standard Bradley-Terry instantiation of DPO is limited in modeling common departures from transitivity in human preferences. To address this, recent work has introduced Self-Play Preference Optimization (SPPO), which iteratively refines the policy by training on self-generated win-lose pairs. Our investigation, however, reveals a critical instability in SPPO: th",
  "authors": "Xiwen Chen, Wenhui Zhu, Jingjing Wang, Peijie Qiu, Zhipeng Wang, Huayu Li et al.",
  "category": "research",
  "topics": "regulation,finance-investment",
  "orgs": null,
  "regions": null,
  "published_at": "2026-06-01T02:06:58.000Z",
  "fetched_at": "2026-07-14T16:30:09.961Z",
  "source_slug": "arxiv-ethics",
  "source_name": "arXiv",
  "source_homepage": "https://arxiv.org",
  "ethics_ai_record_url": "https://ethics.ai/record/3315",
  "original_url": "https://arxiv.org/abs/2606.01561v2",
  "evidence_status": "source-only",
  "attribution": "via ethics.ai"
}