{
  "id": 4494,
  "url": "https://arxiv.org/abs/2605.11712v1",
  "title": "Toward Stable Value Alignment: Introducing Independent Modules for Consistent Value Guidance",
  "summary": "Aligning large language models (LLMs) with human values typically relies on post-training or inference-time steering that directly manipulates the backbone's parameters or representation space. However, a critical gap exists: the model's residual stream is highly dynamic, in which values exist as fragile, low-dimensional properties, inherently incompatible with the stability required for consistent value expression. In this paper, we propose the Stable Value Guidance Transformer (SVGT), which ad",
  "authors": "Wenhao Chen, Sirui Sun, Shengyuan Bai, Guojie Song",
  "category": "research",
  "topics": "safety-alignment",
  "orgs": null,
  "regions": null,
  "published_at": "2026-05-12T08:02:34.000Z",
  "fetched_at": "2026-07-14T16:31:03.578Z",
  "source_slug": "arxiv-ethics",
  "source_name": "arXiv",
  "source_homepage": "https://arxiv.org",
  "ethics_ai_record_url": "https://ethics.ai/record/4494",
  "original_url": "https://arxiv.org/abs/2605.11712v1",
  "evidence_status": "source-only",
  "attribution": "via ethics.ai"
}