{
  "id": 4415,
  "url": "https://arxiv.org/abs/2605.12991v2",
  "title": "Not Just RLHF: Why Alignment Alone Won't Fix Multi-Agent Sycophancy",
  "summary": "LLM-based multi-agent pipelines flip from correct to incorrect answers under simulated peer disagreement at rates we term yield, a vulnerability widely attributed to RLHF-induced sycophancy. We test this attribution across four model families and find it largely wrong: pretrained base models exhibit the same substitution pattern as their Instruct variants, averaging higher yield than Instruct. Using activation patching, we localize the corruption to a narrow mid-layer window where attention carr",
  "authors": "Adarsh Kumarappan, Ananya Mujoo",
  "category": "research",
  "topics": "safety-alignment,agents-autonomy",
  "orgs": null,
  "regions": null,
  "published_at": "2026-05-13T04:45:08.000Z",
  "fetched_at": "2026-07-14T16:30:59.236Z",
  "source_slug": "arxiv-ethics",
  "source_name": "arXiv",
  "source_homepage": "https://arxiv.org",
  "ethics_ai_record_url": "https://ethics.ai/record/4415",
  "original_url": "https://arxiv.org/abs/2605.12991v2",
  "evidence_status": "source-only",
  "attribution": "via ethics.ai"
}