{
  "id": "2610.32246",
  "title": "Do Halting Heads Really Save Compute in Looped Language Models?",
  "authors": "Tsung-Han Wu, Minsuk Heo, Kenny Tan",
  "first_public": "2026-10-23",
  "added_at": "2026-10-26",
  "venue": "arXiv",
  "arxiv": "2610.32246",
  "paper_url": "https://arxiv.org/abs/2610.32246",
  "lens": [
    "recurrent",
    "analysis"
  ],
  "representation": [
    "hidden-state",
    "cache-memory"
  ],
  "computation": [
    "recurrent",
    "depth-wise"
  ],
  "training": [
    "pretraining",
    "analysis-only"
  ],
  "questions": [
    "efficiency",
    "evaluation",
    "limitations"
  ],
  "evidence": [
    "empirical",
    "mechanistic"
  ],
  "modality": [
    "language"
  ],
  "summary": "Compares the stopping probabilities predicted by halting heads with the iterations actually executed. Useful early exits require training every iteration to predict tokens and a strong penalty favoring early stopping; batches that wait for their slowest token erase most of the advantage.",
  "provenance": "Primary-source literature review, 2026-10-26",
  "author_names": [
    "Tsung-Han Wu",
    "Minsuk Heo",
    "Kenny Tan"
  ],
  "record_url": "https://latent-implicit-thinking.github.io/papers/2610.32246/",
  "legacy_record_url": "https://latent-implicit-thinking.github.io/papers/#paper-2610-32246",
  "citation_date": "2026-10-23",
  "json_url": "https://latent-implicit-thinking.github.io/papers/2610.32246/record.json",
  "markdown_url": "https://latent-implicit-thinking.github.io/papers/2610.32246/index.md",
  "bibtex_url": "https://latent-implicit-thinking.github.io/papers/2610.32246/citation.bib",
  "csl_url": "https://latent-implicit-thinking.github.io/papers/2610.32246/citation.csl.json",
  "pdf_url": "https://arxiv.org/pdf/2610.32246"
}