{
  "id": 6418,
  "url": "https://arxiv.org/abs/2604.02652v1",
  "title": "Generalization Limits of Reinforcement Learning Alignment",
  "summary": "The safety of large language models (LLMs) relies on alignment techniques such as reinforcement learning from human feedback (RLHF). However, recent theoretical analyses suggest that reinforcement learning-based training does not acquire new capabilities but merely redistributes the utilization probabilities of existing ones. In this study, we propose ``compound jailbreaks'' targeting OpenAI gpt-oss-20b, which exploit the generalization failures of alignment. This approach combines multiple atta",
  "authors": "Haruhi Shida, Koo Imai, Keigo Kansa",
  "category": "research",
  "topics": "safety-alignment",
  "orgs": "openai",
  "regions": null,
  "published_at": "2026-04-03T02:32:13.000Z",
  "fetched_at": "2026-07-14T16:32:28.610Z",
  "source_slug": "arxiv-ethics",
  "source_name": "arXiv",
  "source_homepage": "https://arxiv.org",
  "ethics_ai_record_url": "https://ethics.ai/record/6418",
  "original_url": "https://arxiv.org/abs/2604.02652v1",
  "evidence_status": "source-only",
  "attribution": "via ethics.ai"
}