{
  "id": "2610.34427",
  "title": "ScaleRLT: Exact Log-Probabilities and Fast Deferred Rollouts for Reinforcement Learning on Looped Transformers",
  "authors": "Dezhen Lu, Zixiang Ni, Runguo Li",
  "first_public": "2026-10-25",
  "added_at": "2026-11-02",
  "venue": "arXiv",
  "arxiv": "2610.34427",
  "paper_url": "https://arxiv.org/abs/2610.34427",
  "html_url": "https://arxiv.org/html/2610.34427v1",
  "lens": [
    "recurrent"
  ],
  "representation": [
    "hidden-state",
    "cache-memory"
  ],
  "computation": [
    "recurrent",
    "depth-wise",
    "parallel"
  ],
  "training": [
    "reinforcement-learning"
  ],
  "questions": [
    "efficiency",
    "architecture"
  ],
  "evidence": [
    "empirical"
  ],
  "modality": [
    "language",
    "symbolic"
  ],
  "summary": "Batches later transformer loops across generated positions to accelerate reinforcement-learning rollouts. Replays the rollout schedule to recover its exact token log probabilities while using a dense pass for gradients, and evaluates the resulting training system on math problems.",
  "provenance": "Primary-source literature review, 2026-11-02",
  "author_names": [
    "Dezhen Lu",
    "Zixiang Ni",
    "Runguo Li"
  ],
  "record_url": "https://latent-implicit-thinking.github.io/papers/2610.34427/",
  "legacy_record_url": "https://latent-implicit-thinking.github.io/papers/#paper-2610-34427",
  "citation_date": "2026-10-25",
  "json_url": "https://latent-implicit-thinking.github.io/papers/2610.34427/record.json",
  "markdown_url": "https://latent-implicit-thinking.github.io/papers/2610.34427/index.md",
  "bibtex_url": "https://latent-implicit-thinking.github.io/papers/2610.34427/citation.bib",
  "csl_url": "https://latent-implicit-thinking.github.io/papers/2610.34427/citation.csl.json",
  "pdf_url": "https://arxiv.org/pdf/2610.34427"
}