{
  "id": 40,
  "url": "https://arxiv.org/abs/2607.10517v1",
  "title": "Conditional Optimal Bridge for Riemannian Activation Steering",
  "summary": "Activation steering offers a lightweight alternative to fine-tuning for controlling large language models at inference time. While many existing methods implicitly optimize a log-density-ratio objective between desired and undesired activation distributions, they do so heuristically rather than deriving it from a principled optimization problem. Moreover, these methods produce query-independent steering directions that can degrade performance on both in-distribution and out-of-distribution (OOD)",
  "authors": "Seyed Arshan Dalili, Ajay Narayanan Sridhar, Vijaykrishnan Narayanan, Mehrdad Mahdavi",
  "category": "research",
  "topics": null,
  "orgs": null,
  "regions": null,
  "published_at": "2026-07-12T00:36:46.000Z",
  "fetched_at": "2026-07-14T14:14:15.664Z",
  "source_slug": "arxiv-ethics",
  "source_name": "arXiv",
  "source_homepage": "https://arxiv.org",
  "ethics_ai_record_url": "https://ethics.ai/record/40",
  "original_url": "https://arxiv.org/abs/2607.10517v1",
  "evidence_status": "source-only",
  "attribution": "via ethics.ai"
}