{
  "id": 4798,
  "url": "https://arxiv.org/abs/2605.06908v1",
  "title": "Same Signal, Opposite Meaning: Direction-Informed Adaptive Learning for LLM Agents",
  "summary": "Adaptive test-time compute for LLM agents aims to invoke extra computation only when it improves performance. Existing methods typically use confidence-, uncertainty-, or difficulty-based gates, assuming a fixed direction from the gating signal through compute need to the value of computation. This makes gating a utility-calibration problem: gating signals should align with whether extra computation improves the final outcome over the base policy. We show that this alignment is unstable: the sam",
  "authors": "Ziming Li, Jiatan Huang, Xiaoguang Guo, Guilin Wang, Chuxu Zhang",
  "category": "research",
  "topics": "regulation,safety-alignment,agents-autonomy",
  "orgs": null,
  "regions": null,
  "published_at": "2026-05-07T20:15:51.000Z",
  "fetched_at": "2026-07-14T16:31:17.581Z",
  "source_slug": "arxiv-ethics",
  "source_name": "arXiv",
  "source_homepage": "https://arxiv.org",
  "ethics_ai_record_url": "https://ethics.ai/record/4798",
  "original_url": "https://arxiv.org/abs/2605.06908v1",
  "evidence_status": "source-only",
  "attribution": "via ethics.ai"
}