{
  "id": 4943,
  "url": "https://arxiv.org/abs/2605.04330v1",
  "title": "The Scaling Properties of Implicit Deductive Reasoning in Transformers",
  "summary": "We investigate the scaling properties of implicit deductive reasoning over Horn clauses in depth-bounded Transformers. By systematically decorrelating provability from spurious features and enforcing algorithmic alignment, we find that in sufficiently deep models with a bidirectional prefix mask, implicit reasoning approaches explicit CoT performance across graph topologies and problem widths, though CoT remains necessary for depth extrapolation.",
  "authors": "Enrico Vompa, Tanel Tammet",
  "category": "research",
  "topics": "safety-alignment,finance-investment",
  "orgs": null,
  "regions": null,
  "published_at": "2026-05-05T22:35:46.000Z",
  "fetched_at": "2026-07-14T16:31:21.935Z",
  "source_slug": "arxiv-ethics",
  "source_name": "arXiv",
  "source_homepage": "https://arxiv.org",
  "ethics_ai_record_url": "https://ethics.ai/record/4943",
  "original_url": "https://arxiv.org/abs/2605.04330v1",
  "evidence_status": "source-only",
  "attribution": "via ethics.ai"
}