{
  "id": 3557,
  "url": "https://arxiv.org/abs/2605.29123v1",
  "title": "The Confidence Shortcut: A Reasoning Failure Mode of Masked Diffusion Models",
  "summary": "Masked diffusion language models (MDMs) uniquely support any-order generation, with confidence-based decoding currently serving as the de facto standard inference policy. To optimize for this, recent training schemes attempt to align training mask patterns directly with those observed during generation. However, we argue that confidence-based decoding is inherently misaligned with the logical-flow trajectories required for complex reasoning, and that confidence-aligned training actively entrench",
  "authors": "Dueun Kim, Albert No",
  "category": "research",
  "topics": "regulation,safety-alignment",
  "orgs": null,
  "regions": null,
  "published_at": "2026-05-27T21:33:37.000Z",
  "fetched_at": "2026-07-14T16:30:18.858Z",
  "source_slug": "arxiv-ethics",
  "source_name": "arXiv",
  "source_homepage": "https://arxiv.org",
  "ethics_ai_record_url": "https://ethics.ai/record/3557",
  "original_url": "https://arxiv.org/abs/2605.29123v1",
  "evidence_status": "source-only",
  "attribution": "via ethics.ai"
}