{
  "id": "2412.06769",
  "title": "Training Large Language Models to Reason in a Continuous Latent Space",
  "authors": "Shibo Hao, Sainbayar Sukhbaatar, DiJia Su, Xian Li, Zhiting Hu, Jason Weston, Yuandong Tian",
  "first_public": "2024-12-09",
  "venue": "COLM 2025",
  "arxiv": "2412.06769",
  "paper_url": "https://arxiv.org/abs/2412.06769",
  "code_url": "https://github.com/facebookresearch/coconut",
  "lens": [
    "continuous"
  ],
  "representation": [
    "continuous-state"
  ],
  "computation": [
    "recurrent",
    "token-wise"
  ],
  "training": [
    "curriculum"
  ],
  "questions": [
    "capability",
    "efficiency"
  ],
  "evidence": [
    "empirical"
  ],
  "modality": [
    "language",
    "symbolic"
  ],
  "summary": "Feeds a model's last hidden state back as its next input, enabling recurrent reasoning in continuous space before returning to language.",
  "provenance": "Awesome-Latent-CoT",
  "author_names": [
    "Shibo Hao",
    "Sainbayar Sukhbaatar",
    "DiJia Su",
    "Xian Li",
    "Zhiting Hu",
    "Jason Weston",
    "Yuandong Tian"
  ],
  "record_url": "https://latent-implicit-thinking.github.io/papers/2412.06769/",
  "legacy_record_url": "https://latent-implicit-thinking.github.io/papers/#paper-2412-06769",
  "citation_date": "2024-12-09",
  "json_url": "https://latent-implicit-thinking.github.io/papers/2412.06769/record.json",
  "markdown_url": "https://latent-implicit-thinking.github.io/papers/2412.06769/index.md",
  "bibtex_url": "https://latent-implicit-thinking.github.io/papers/2412.06769/citation.bib",
  "csl_url": "https://latent-implicit-thinking.github.io/papers/2412.06769/citation.csl.json",
  "publication": {
    "title": "Training Large Language Models to Reason in a Continuous Latent Space",
    "author_names": [
      "Shibo Hao",
      "Sainbayar Sukhbaatar",
      "DiJia Su",
      "Xian Li",
      "Zhiting Hu",
      "Jason E Weston",
      "Yuandong Tian"
    ],
    "year": 2025,
    "type": "paper-conference",
    "container_title": "Conference on Language Modeling (COLM 2025)",
    "url": "https://openreview.net/forum?id=Itxz7S4Ip3",
    "source_url": "https://colmweb.org/2025/AcceptedPapers.html",
    "verified_at": "2026-10-18"
  },
  "publication_bibtex_url": "https://latent-implicit-thinking.github.io/papers/2412.06769/publication.bib",
  "publication_csl_url": "https://latent-implicit-thinking.github.io/papers/2412.06769/publication.csl.json",
  "method_comparisons": [
    {
      "title": "Implicit CoT: proposal and implementation",
      "text": "Deng et al. (2023) proposed implicit chain-of-thought reasoning: language models carry out intermediate reasoning in continuous hidden states rather than generate it as text. Section 2.3 defines this setting; section 3 develops a distillation method to implement it. That method transfers a teacher's reasoning into hidden states across layers. Reasoning through continuous hidden states is the broader proposal; it does not require the particular distillation procedure used in this paper. Its across-layer computation differs from the repeated hidden-state feedback used by later methods such as Coconut.",
      "source_ids": [
        "2311.01460",
        "2412.06769"
      ],
      "source_urls": [
        "https://arxiv.org/html/2311.01460v1#S2.SS3",
        "https://arxiv.org/html/2311.01460v1#S3",
        "https://arxiv.org/html/2412.06769v3#S3"
      ],
      "source_labels": [
        "Implicit CoT proposal, §2.3",
        "Distillation implementation, §3",
        "Coconut, §3"
      ],
      "topic_id": "continuous",
      "topic_label": "Continuous-state reasoning",
      "topic_url": "https://latent-implicit-thinking.github.io/topics/continuous-states/"
    },
    {
      "title": "From stepwise internalization to Coconut",
      "text": "Coconut explicitly credits Deng et al. (2024) as the inspiration for its multistage training curriculum. Stepwise internalization gradually removes written CoT tokens until the model answers directly. Coconut instead replaces removed steps with continuous hidden-state inputs, retaining a sequence of internal reasoning steps before the answer. It also follows the earlier work in resetting the optimizer between training stages.",
      "source_ids": [
        "2405.14838",
        "2412.06769"
      ],
      "source_urls": [
        "https://arxiv.org/html/2405.14838v1#S3",
        "https://arxiv.org/html/2412.06769v3#S3"
      ],
      "source_labels": [
        "Deng et al. (2024), §3",
        "Coconut, §3"
      ],
      "topic_id": "continuous",
      "topic_label": "Continuous-state reasoning",
      "topic_url": "https://latent-implicit-thinking.github.io/topics/continuous-states/"
    },
    {
      "title": "Coconut and CODI: similar computation, different training",
      "text": "Coconut and CODI both pass continuous states between reasoning steps, but supervise them differently. Coconut gradually replaces written steps with hidden-state inputs and trains on the remaining text. CODI jointly trains explicit and implicit reasoning tasks, matching their hidden states at a selected position before the answer. These models avoid decoding a word at every intermediate step; they still perform successive model passes, so fewer written tokens does not mean no reasoning-time computation.",
      "source_ids": [
        "2412.06769",
        "2502.21074"
      ],
      "source_urls": [
        "https://arxiv.org/html/2412.06769v3#S3",
        "https://arxiv.org/html/2502.21074v3#S3"
      ],
      "source_labels": [
        "Coconut, §3",
        "CODI, §3"
      ],
      "topic_id": "continuous",
      "topic_label": "Continuous-state reasoning",
      "topic_url": "https://latent-implicit-thinking.github.io/topics/continuous-states/"
    }
  ]
}