{
  "id": "2602.22441",
  "title": "How Do Latent Reasoning Methods Perform Under Weak and Strong Supervision?",
  "authors": "Yingqian Cui, Zhenwei Dai, Bing He, Zhan Shi, Hui Liu, Rui Sun, Zhiji Liu, Yue Xing, Jiliang Tang, Benoit Dumoulin",
  "first_public": "2026-02-25",
  "added_at": "2026-11-03",
  "venue": "arXiv",
  "arxiv": "2602.22441",
  "paper_url": "https://arxiv.org/abs/2602.22441",
  "html_url": "https://arxiv.org/html/2602.22441v1",
  "lens": [
    "analysis",
    "continuous"
  ],
  "representation": [
    "hidden-state",
    "text-latent-hybrid"
  ],
  "computation": [
    "token-wise"
  ],
  "training": [
    "analysis-only",
    "supervised"
  ],
  "questions": [
    "interpretability",
    "evaluation"
  ],
  "evidence": [
    "empirical",
    "mechanistic"
  ],
  "modality": [
    "language",
    "symbolic"
  ],
  "summary": "Tests latent reasoning on GSM8K and ProsQA by removing steps, perturbing hidden states, and sampling continuations. Finds that stronger supervision reduces shortcut use but narrows sampled solutions, and studies changes to Coconut training that increase its reliance on latent steps.",
  "provenance": "Primary-source literature review, 2026-11-03",
  "author_names": [
    "Yingqian Cui",
    "Zhenwei Dai",
    "Bing He",
    "Zhan Shi",
    "Hui Liu",
    "Rui Sun",
    "Zhiji Liu",
    "Yue Xing",
    "Jiliang Tang",
    "Benoit Dumoulin"
  ],
  "record_url": "https://latent-implicit-thinking.github.io/papers/2602.22441/",
  "legacy_record_url": "https://latent-implicit-thinking.github.io/papers/#paper-2602-22441",
  "citation_date": "2026-02-25",
  "json_url": "https://latent-implicit-thinking.github.io/papers/2602.22441/record.json",
  "markdown_url": "https://latent-implicit-thinking.github.io/papers/2602.22441/index.md",
  "bibtex_url": "https://latent-implicit-thinking.github.io/papers/2602.22441/citation.bib",
  "csl_url": "https://latent-implicit-thinking.github.io/papers/2602.22441/citation.csl.json",
  "pdf_url": "https://arxiv.org/pdf/2602.22441"
}