{
  "id": "2510.10089",
  "title": "What Makes Looped Transformers Perform Better Than Non-Recursive Ones",
  "authors": "Zixuan Gong, Yong Liu, Jiaye Teng",
  "first_public": "2025-10-11",
  "added_at": "2026-11-03",
  "venue": "arXiv",
  "arxiv": "2510.10089",
  "paper_url": "https://arxiv.org/abs/2510.10089",
  "html_url": "https://arxiv.org/html/2510.10089v3",
  "lens": [
    "recurrent",
    "analysis"
  ],
  "representation": [
    "hidden-state"
  ],
  "computation": [
    "recurrent",
    "depth-wise"
  ],
  "training": [
    "supervised",
    "curriculum"
  ],
  "questions": [
    "theory",
    "efficiency",
    "generalization"
  ],
  "evidence": [
    "formal",
    "empirical"
  ],
  "modality": [
    "language",
    "symbolic"
  ],
  "summary": "Studies how repeated attention changes optimization through a model of loss-landscape geometry. Proposes SHIFT, which starts with one attention pass and adds shared-weight repetitions after learning stalls, and tests the schedule on synthetic sequence tasks.",
  "provenance": "Primary-source literature review, 2026-11-03",
  "previous_titles": [
    "What Makes Looped Transformers Perform Better Than Non-Recursive Ones (Provably)"
  ],
  "author_names": [
    "Zixuan Gong",
    "Yong Liu",
    "Jiaye Teng"
  ],
  "record_url": "https://latent-implicit-thinking.github.io/papers/2510.10089/",
  "legacy_record_url": "https://latent-implicit-thinking.github.io/papers/#paper-2510-10089",
  "citation_date": "2025-10-11",
  "json_url": "https://latent-implicit-thinking.github.io/papers/2510.10089/record.json",
  "markdown_url": "https://latent-implicit-thinking.github.io/papers/2510.10089/index.md",
  "bibtex_url": "https://latent-implicit-thinking.github.io/papers/2510.10089/citation.bib",
  "csl_url": "https://latent-implicit-thinking.github.io/papers/2510.10089/citation.csl.json",
  "pdf_url": "https://arxiv.org/pdf/2510.10089"
}