{
  "id": "2610.09013",
  "title": "Enabling Dynamic Computation in Looped LMs",
  "authors": "Aayush Mishra, Arnau Padrés Masdemont, Victor Conchello Vendrell, Jordi Ros Giralt, Arash Behboodi, Fabio Valerio Massoli",
  "first_public": "2026-10-06",
  "added_at": "2026-10-22",
  "venue": "arXiv",
  "arxiv": "2610.09013",
  "paper_url": "https://arxiv.org/abs/2610.09013",
  "html_url": "https://arxiv.org/html/2610.09013v1",
  "lens": [
    "recurrent"
  ],
  "representation": [
    "hidden-state",
    "cache-memory"
  ],
  "computation": [
    "recurrent",
    "depth-wise"
  ],
  "training": [
    "inference-only",
    "pretraining"
  ],
  "questions": [
    "architecture",
    "efficiency"
  ],
  "evidence": [
    "empirical"
  ],
  "modality": [
    "language"
  ],
  "summary": "Lets tokens leave a looped language model at different depths by reusing their final key-value cache entries in later loops. Adds an exit-depth training objective and ragged-cache training to make variable-depth computation practical.",
  "provenance": "Primary-source literature review, 2026-10-22",
  "author_names": [
    "Aayush Mishra",
    "Arnau Padrés Masdemont",
    "Victor Conchello Vendrell",
    "Jordi Ros Giralt",
    "Arash Behboodi",
    "Fabio Valerio Massoli"
  ],
  "record_url": "https://latent-implicit-thinking.github.io/papers/2610.09013/",
  "legacy_record_url": "https://latent-implicit-thinking.github.io/papers/#paper-2610-09013",
  "citation_date": "2026-10-06",
  "json_url": "https://latent-implicit-thinking.github.io/papers/2610.09013/record.json",
  "markdown_url": "https://latent-implicit-thinking.github.io/papers/2610.09013/index.md",
  "bibtex_url": "https://latent-implicit-thinking.github.io/papers/2610.09013/citation.bib",
  "csl_url": "https://latent-implicit-thinking.github.io/papers/2610.09013/citation.csl.json",
  "pdf_url": "https://arxiv.org/pdf/2610.09013"
}