{
  "id": 5583,
  "url": "https://arxiv.org/abs/2604.18789v1",
  "title": "ARES: Adaptive Red-Teaming and End-to-End Repair of Policy-Reward System",
  "summary": "Reinforcement Learning from Human Feedback (RLHF) is central to aligning Large Language Models (LLMs), yet it introduces a critical vulnerability: an imperfect Reward Model (RM) can become a single point of failure when it fails to penalize unsafe behaviors. While existing red-teaming approaches primarily target policy-level weaknesses, they overlook what we term systemic weaknesses cases where both the core LLM and the RM fail in tandem. We present ARES, a framework that systematically discover",
  "authors": "Jiacheng Liang, Yao Ma, Tharindu Kumarage, Satyapriya Krishna, Rahul Gupta, Kai-Wei Chang et al.",
  "category": "research",
  "topics": "regulation,safety-alignment",
  "orgs": null,
  "regions": null,
  "published_at": "2026-04-20T19:54:47.000Z",
  "fetched_at": "2026-07-14T16:31:53.163Z",
  "source_slug": "arxiv-ethics",
  "source_name": "arXiv",
  "source_homepage": "https://arxiv.org",
  "ethics_ai_record_url": "https://ethics.ai/record/5583",
  "original_url": "https://arxiv.org/abs/2604.18789v1",
  "evidence_status": "source-only",
  "attribution": "via ethics.ai"
}