{
  "metadata": {
    "version": "0.9.4",
    "updated_at": "2026-10-18",
    "search_refreshed_at": "2026-10-18",
    "topic_id": "learned-tokens",
    "name": "Thinking tokens",
    "definition": "Give a model extra tokens for computation, such as pauses, filler symbols, or learned planning tokens. These tokens need not spell out the solution.",
    "record_count": 18,
    "topic_url": "https://latent-implicit-thinking.github.io/topics/thinking-tokens/",
    "source_url": "https://latent-implicit-thinking.github.io/data/papers.json"
  },
  "comparisons": [
    {
      "title": "Pausing and planning",
      "text": "Pause tokens add computation before the answer without choosing a different plan token for each step. Planning tokens instead select a learned token before generating each written reasoning step. The latter still produces a textual chain of thought. In the pause-token experiments, the training setup matters: gains were strongest when pauses were used in both pretraining and fine-tuning, not simply inserted into an unchanged model.",
      "source_ids": [
        "2310.02226",
        "2310.05707"
      ],
      "source_urls": [
        "https://proceedings.iclr.cc/paper_files/paper/2024/hash/76917808731dae9e6d62c2a7a6afb542-Abstract-Conference.html",
        "https://arxiv.org/html/2310.05707v3#S2"
      ],
      "source_labels": [
        "Pause-token paper",
        "Planning-token paper, §2"
      ]
    }
  ],
  "demos": [

  ],
  "papers": [
    {
      "id": "2405.08644",
      "title": "Thinking Tokens for Language Modeling",
      "authors": "David Herel, Tomas Mikolov",
      "first_public": "2023-09-04",
      "first_public_source": "https://aitp-conference.org/2023/",
      "first_public_note": "AITP presentation date; the earlier public release date of the conference abstract is unverified.",
      "citation_date": "2024-05-14",
      "venue": "AITP 2023",
      "arxiv": "2405.08644",
      "paper_url": "https://arxiv.org/abs/2405.08644",
      "lens": [
        "learned-tokens"
      ],
      "representation": [
        "discrete-token"
      ],
      "computation": [
        "token-wise"
      ],
      "training": [
        "pretraining"
      ],
      "questions": [
        "capability",
        "efficiency"
      ],
      "evidence": [
        "empirical"
      ],
      "modality": [
        "language"
      ],
      "summary": "Gives a recurrent language model extra hidden-state updates by inserting special tokens after each word. Tests whether these updates improve next-word prediction.",
      "provenance": "Primary-source literature search, 2026-09-04",
      "author_names": [
        "David Herel",
        "Tomas Mikolov"
      ],
      "record_url": "https://latent-implicit-thinking.github.io/papers/2405.08644/",
      "legacy_record_url": "https://latent-implicit-thinking.github.io/papers/#paper-2405-08644",
      "json_url": "https://latent-implicit-thinking.github.io/papers/2405.08644/record.json",
      "markdown_url": "https://latent-implicit-thinking.github.io/papers/2405.08644/index.md",
      "bibtex_url": "https://latent-implicit-thinking.github.io/papers/2405.08644/citation.bib",
      "csl_url": "https://latent-implicit-thinking.github.io/papers/2405.08644/citation.csl.json"
    },
    {
      "id": "2310.02226",
      "title": "Think before you speak: Training language models with pause tokens",
      "authors": "Sachin Goyal, Ziwei Ji, Ankit Singh Rawat, Aditya Krishna Menon, Sanjiv Kumar, Vaishnavh Nagarajan",
      "first_public": "2023-10-03",
      "venue": "ICLR 2024",
      "arxiv": "2310.02226",
      "paper_url": "https://arxiv.org/abs/2310.02226",
      "lens": [
        "learned-tokens"
      ],
      "representation": [
        "discrete-token"
      ],
      "computation": [
        "token-wise"
      ],
      "training": [
        "pretraining"
      ],
      "questions": [
        "capability"
      ],
      "evidence": [
        "empirical"
      ],
      "modality": [
        "language"
      ],
      "summary": "Introduces pause tokens that give a language model additional computational positions before it must produce an answer.",
      "provenance": "Awesome-Latent-CoT",
      "author_names": [
        "Sachin Goyal",
        "Ziwei Ji",
        "Ankit Singh Rawat",
        "Aditya Krishna Menon",
        "Sanjiv Kumar",
        "Vaishnavh Nagarajan"
      ],
      "record_url": "https://latent-implicit-thinking.github.io/papers/2310.02226/",
      "legacy_record_url": "https://latent-implicit-thinking.github.io/papers/#paper-2310-02226",
      "citation_date": "2023-10-03",
      "json_url": "https://latent-implicit-thinking.github.io/papers/2310.02226/record.json",
      "markdown_url": "https://latent-implicit-thinking.github.io/papers/2310.02226/index.md",
      "bibtex_url": "https://latent-implicit-thinking.github.io/papers/2310.02226/citation.bib",
      "csl_url": "https://latent-implicit-thinking.github.io/papers/2310.02226/citation.csl.json",
      "publication": {
        "title": "Think before you speak: Training Language Models With Pause Tokens",
        "author_names": [
          "Sachin Goyal",
          "Ziwei Ji",
          "Ankit Singh Rawat",
          "Aditya Krishna Menon",
          "Sanjiv Kumar",
          "Vaishnavh Nagarajan"
        ],
        "year": 2024,
        "type": "paper-conference",
        "container_title": "International Conference on Learning Representations",
        "pages": "27896-27923",
        "volume": "2024",
        "url": "https://proceedings.iclr.cc/paper_files/paper/2024/hash/76917808731dae9e6d62c2a7a6afb542-Abstract-Conference.html",
        "source_url": "https://proceedings.iclr.cc/paper_files/paper/2024/hash/76917808731dae9e6d62c2a7a6afb542-Abstract-Conference.html",
        "verified_at": "2026-10-18"
      },
      "publication_bibtex_url": "https://latent-implicit-thinking.github.io/papers/2310.02226/publication.bib",
      "publication_csl_url": "https://latent-implicit-thinking.github.io/papers/2310.02226/publication.csl.json",
      "method_comparisons": [
        {
          "title": "Pausing and planning",
          "text": "Pause tokens add computation before the answer without choosing a different plan token for each step. Planning tokens instead select a learned token before generating each written reasoning step. The latter still produces a textual chain of thought. In the pause-token experiments, the training setup matters: gains were strongest when pauses were used in both pretraining and fine-tuning, not simply inserted into an unchanged model.",
          "source_ids": [
            "2310.02226",
            "2310.05707"
          ],
          "source_urls": [
            "https://proceedings.iclr.cc/paper_files/paper/2024/hash/76917808731dae9e6d62c2a7a6afb542-Abstract-Conference.html",
            "https://arxiv.org/html/2310.05707v3#S2"
          ],
          "source_labels": [
            "Pause-token paper",
            "Planning-token paper, §2"
          ],
          "topic_id": "learned-tokens",
          "topic_label": "Thinking tokens",
          "topic_url": "https://latent-implicit-thinking.github.io/topics/thinking-tokens/"
        }
      ]
    },
    {
      "id": "2310.05707",
      "title": "Guiding Language Model Reasoning with Planning Tokens",
      "authors": "Xinyi Wang, Lucas Caccia, Oleksiy Ostapenko, Xingdi Yuan, William Yang Wang, Alessandro Sordoni",
      "first_public": "2023-10-09",
      "venue": "COLM 2024",
      "arxiv": "2310.05707",
      "paper_url": "https://arxiv.org/abs/2310.05707",
      "code_url": "https://github.com/WANGXinyiLinda/planning_tokens",
      "lens": [
        "learned-tokens"
      ],
      "representation": [
        "discrete-token"
      ],
      "computation": [
        "token-wise"
      ],
      "training": [
        "supervised"
      ],
      "questions": [
        "capability",
        "controllability"
      ],
      "evidence": [
        "empirical"
      ],
      "modality": [
        "language"
      ],
      "summary": "Adds a learned planning token before each written reasoning step, then uses that token to guide the step.",
      "provenance": "Awesome-Latent-CoT",
      "author_names": [
        "Xinyi Wang",
        "Lucas Caccia",
        "Oleksiy Ostapenko",
        "Xingdi Yuan",
        "William Yang Wang",
        "Alessandro Sordoni"
      ],
      "record_url": "https://latent-implicit-thinking.github.io/papers/2310.05707/",
      "legacy_record_url": "https://latent-implicit-thinking.github.io/papers/#paper-2310-05707",
      "citation_date": "2023-10-09",
      "json_url": "https://latent-implicit-thinking.github.io/papers/2310.05707/record.json",
      "markdown_url": "https://latent-implicit-thinking.github.io/papers/2310.05707/index.md",
      "bibtex_url": "https://latent-implicit-thinking.github.io/papers/2310.05707/citation.bib",
      "csl_url": "https://latent-implicit-thinking.github.io/papers/2310.05707/citation.csl.json",
      "method_comparisons": [
        {
          "title": "Pausing and planning",
          "text": "Pause tokens add computation before the answer without choosing a different plan token for each step. Planning tokens instead select a learned token before generating each written reasoning step. The latter still produces a textual chain of thought. In the pause-token experiments, the training setup matters: gains were strongest when pauses were used in both pretraining and fine-tuning, not simply inserted into an unchanged model.",
          "source_ids": [
            "2310.02226",
            "2310.05707"
          ],
          "source_urls": [
            "https://proceedings.iclr.cc/paper_files/paper/2024/hash/76917808731dae9e6d62c2a7a6afb542-Abstract-Conference.html",
            "https://arxiv.org/html/2310.05707v3#S2"
          ],
          "source_labels": [
            "Pause-token paper",
            "Planning-token paper, §2"
          ],
          "topic_id": "learned-tokens",
          "topic_label": "Thinking tokens",
          "topic_url": "https://latent-implicit-thinking.github.io/topics/thinking-tokens/"
        }
      ]
    },
    {
      "id": "2403.09629",
      "title": "Quiet-STaR: Language Models Can Teach Themselves to Think Before Speaking",
      "authors": "Eric Zelikman, Georges Harik, Yijia Shao, Varuna Jayasiri, Nick Haber, Noah D. Goodman",
      "first_public": "2024-03-14",
      "venue": "COLM 2024",
      "arxiv": "2403.09629",
      "paper_url": "https://arxiv.org/abs/2403.09629",
      "lens": [
        "learned-tokens"
      ],
      "representation": [
        "text-latent-hybrid"
      ],
      "computation": [
        "token-wise"
      ],
      "training": [
        "self-training"
      ],
      "questions": [
        "capability"
      ],
      "evidence": [
        "empirical"
      ],
      "modality": [
        "language"
      ],
      "summary": "Trains on ordinary text by generating written rationales before predicting what comes next. Learns from how much those rationales improve subsequent token predictions.",
      "provenance": "Awesome-Latent-CoT",
      "author_names": [
        "Eric Zelikman",
        "Georges Harik",
        "Yijia Shao",
        "Varuna Jayasiri",
        "Nick Haber",
        "Noah D. Goodman"
      ],
      "record_url": "https://latent-implicit-thinking.github.io/papers/2403.09629/",
      "legacy_record_url": "https://latent-implicit-thinking.github.io/papers/#paper-2403-09629",
      "citation_date": "2024-03-14",
      "json_url": "https://latent-implicit-thinking.github.io/papers/2403.09629/record.json",
      "markdown_url": "https://latent-implicit-thinking.github.io/papers/2403.09629/index.md",
      "bibtex_url": "https://latent-implicit-thinking.github.io/papers/2403.09629/citation.bib",
      "csl_url": "https://latent-implicit-thinking.github.io/papers/2403.09629/citation.csl.json"
    },
    {
      "id": "2404.15758",
      "title": "Let's think dot by dot: Hidden computation in transformer language models",
      "authors": "Jacob Pfau, William Merrill, Samuel R. Bowman",
      "first_public": "2024-04-24",
      "venue": "COLM 2024",
      "arxiv": "2404.15758",
      "paper_url": "https://arxiv.org/abs/2404.15758",
      "code_url": "https://github.com/JacobPfau/fillerTokens",
      "lens": [
        "learned-tokens"
      ],
      "representation": [
        "discrete-token"
      ],
      "computation": [
        "token-wise"
      ],
      "training": [
        "supervised"
      ],
      "questions": [
        "capability",
        "theory"
      ],
      "evidence": [
        "empirical"
      ],
      "modality": [
        "language",
        "symbolic"
      ],
      "summary": "Studies whether meaningless filler-token positions can provide transformers with additional hidden computation without carrying readable reasoning content.",
      "provenance": "Awesome-Latent-CoT",
      "author_names": [
        "Jacob Pfau",
        "William Merrill",
        "Samuel R. Bowman"
      ],
      "record_url": "https://latent-implicit-thinking.github.io/papers/2404.15758/",
      "legacy_record_url": "https://latent-implicit-thinking.github.io/papers/#paper-2404-15758",
      "citation_date": "2024-04-24",
      "json_url": "https://latent-implicit-thinking.github.io/papers/2404.15758/record.json",
      "markdown_url": "https://latent-implicit-thinking.github.io/papers/2404.15758/index.md",
      "bibtex_url": "https://latent-implicit-thinking.github.io/papers/2404.15758/citation.bib",
      "csl_url": "https://latent-implicit-thinking.github.io/papers/2404.15758/citation.csl.json"
    },
    {
      "id": "2502.03275",
      "title": "Token Assorted: Mixing Latent and Text Tokens for Improved Language Model Reasoning",
      "authors": "DiJia Su, Hanlin Zhu, Yingchen Xu, Jiantao Jiao, Yuandong Tian, Qinqing Zheng",
      "first_public": "2025-02-05",
      "venue": "arXiv",
      "arxiv": "2502.03275",
      "paper_url": "https://arxiv.org/abs/2502.03275",
      "lens": [
        "learned-tokens",
        "continuous"
      ],
      "representation": [
        "discrete-token",
        "text-latent-hybrid"
      ],
      "computation": [
        "token-wise"
      ],
      "training": [
        "supervised"
      ],
      "questions": [
        "capability",
        "efficiency"
      ],
      "evidence": [
        "empirical"
      ],
      "modality": [
        "language",
        "symbolic"
      ],
      "summary": "Interleaves latent and textual tokens so a model can combine compressed internal computation with readable reasoning steps.",
      "provenance": "Primary-source literature search, 2026-09-04",
      "author_names": [
        "DiJia Su",
        "Hanlin Zhu",
        "Yingchen Xu",
        "Jiantao Jiao",
        "Yuandong Tian",
        "Qinqing Zheng"
      ],
      "record_url": "https://latent-implicit-thinking.github.io/papers/2502.03275/",
      "legacy_record_url": "https://latent-implicit-thinking.github.io/papers/#paper-2502-03275",
      "citation_date": "2025-02-05",
      "json_url": "https://latent-implicit-thinking.github.io/papers/2502.03275/record.json",
      "markdown_url": "https://latent-implicit-thinking.github.io/papers/2502.03275/index.md",
      "bibtex_url": "https://latent-implicit-thinking.github.io/papers/2502.03275/citation.bib",
      "csl_url": "https://latent-implicit-thinking.github.io/papers/2502.03275/citation.csl.json"
    },
    {
      "id": "2504.14992",
      "title": "Efficient Pretraining Length Scaling",
      "authors": "Bohong Wu, Shen Yan, Sijun Zhang, Jianqiao Lu, Yutao Zeng, Ya Wang, Xun Zhou",
      "first_public": "2025-04-21",
      "venue": "arXiv",
      "arxiv": "2504.14992",
      "paper_url": "https://arxiv.org/abs/2504.14992",
      "lens": [
        "learned-tokens",
        "recurrent"
      ],
      "representation": [
        "hidden-state",
        "cache-memory"
      ],
      "computation": [
        "token-wise",
        "stateful"
      ],
      "training": [
        "pretraining"
      ],
      "questions": [
        "efficiency",
        "scaling",
        "architecture"
      ],
      "evidence": [
        "empirical",
        "scaling"
      ],
      "modality": [
        "language"
      ],
      "summary": "Adds hidden decoding tokens during pretraining, then discards their key/value cache entries after use. Keeps the original tokens' entries to preserve longer-range context.",
      "provenance": "Primary-source literature search, 2026-09-04",
      "author_names": [
        "Bohong Wu",
        "Shen Yan",
        "Sijun Zhang",
        "Jianqiao Lu",
        "Yutao Zeng",
        "Ya Wang",
        "Xun Zhou"
      ],
      "record_url": "https://latent-implicit-thinking.github.io/papers/2504.14992/",
      "legacy_record_url": "https://latent-implicit-thinking.github.io/papers/#paper-2504-14992",
      "citation_date": "2025-04-21",
      "json_url": "https://latent-implicit-thinking.github.io/papers/2504.14992/record.json",
      "markdown_url": "https://latent-implicit-thinking.github.io/papers/2504.14992/index.md",
      "bibtex_url": "https://latent-implicit-thinking.github.io/papers/2504.14992/citation.bib",
      "csl_url": "https://latent-implicit-thinking.github.io/papers/2504.14992/citation.csl.json"
    },
    {
      "id": "2505.12629",
      "title": "Enhancing Latent Computation in Transformers with Latent Tokens",
      "authors": "Yuchang Sun, Yanxi Chen, Yaliang Li, Bolin Ding",
      "first_public": "2025-05-19",
      "venue": "arXiv",
      "arxiv": "2505.12629",
      "paper_url": "https://arxiv.org/abs/2505.12629",
      "lens": [
        "learned-tokens"
      ],
      "representation": [
        "discrete-token"
      ],
      "computation": [
        "token-wise"
      ],
      "training": [
        "supervised"
      ],
      "questions": [
        "capability",
        "efficiency"
      ],
      "evidence": [
        "empirical"
      ],
      "modality": [
        "language",
        "symbolic"
      ],
      "summary": "Adds trainable dummy tokens to a pretrained transformer so attention can use them for extra computation during decoding. The tokens can be inserted in different positions at inference time.",
      "provenance": "Primary-source literature search, 2026-09-04",
      "author_names": [
        "Yuchang Sun",
        "Yanxi Chen",
        "Yaliang Li",
        "Bolin Ding"
      ],
      "record_url": "https://latent-implicit-thinking.github.io/papers/2505.12629/",
      "legacy_record_url": "https://latent-implicit-thinking.github.io/papers/#paper-2505-12629",
      "citation_date": "2025-05-19",
      "json_url": "https://latent-implicit-thinking.github.io/papers/2505.12629/record.json",
      "markdown_url": "https://latent-implicit-thinking.github.io/papers/2505.12629/index.md",
      "bibtex_url": "https://latent-implicit-thinking.github.io/papers/2505.12629/citation.bib",
      "csl_url": "https://latent-implicit-thinking.github.io/papers/2505.12629/citation.csl.json"
    },
    {
      "id": "2506.17218",
      "title": "Machine Mental Imagery: Empower Multimodal Reasoning with Latent Visual Tokens",
      "authors": "Zeyuan Yang, Xueyang Yu, Delin Chen, Maohao Shen, Chuang Gan",
      "first_public": "2025-06-20",
      "venue": "arXiv",
      "arxiv": "2506.17218",
      "paper_url": "https://arxiv.org/abs/2506.17218",
      "lens": [
        "learned-tokens",
        "continuous"
      ],
      "representation": [
        "multimodal-latent"
      ],
      "computation": [
        "token-wise"
      ],
      "training": [
        "supervised"
      ],
      "questions": [
        "capability",
        "interpretability"
      ],
      "evidence": [
        "empirical"
      ],
      "modality": [
        "multimodal",
        "vision"
      ],
      "summary": "Uses latent visual tokens as an internal imagery channel for multimodal reasoning without rendering every intermediate state.",
      "provenance": "Primary-source literature search, 2026-09-04",
      "author_names": [
        "Zeyuan Yang",
        "Xueyang Yu",
        "Delin Chen",
        "Maohao Shen",
        "Chuang Gan"
      ],
      "record_url": "https://latent-implicit-thinking.github.io/papers/2506.17218/",
      "legacy_record_url": "https://latent-implicit-thinking.github.io/papers/#paper-2506-17218",
      "citation_date": "2025-06-20",
      "json_url": "https://latent-implicit-thinking.github.io/papers/2506.17218/record.json",
      "markdown_url": "https://latent-implicit-thinking.github.io/papers/2506.17218/index.md",
      "bibtex_url": "https://latent-implicit-thinking.github.io/papers/2506.17218/citation.bib",
      "csl_url": "https://latent-implicit-thinking.github.io/papers/2506.17218/citation.csl.json"
    },
    {
      "id": "2601.21576",
      "title": "Chain Of Thought Compression: A Theoretical Analysis",
      "authors": "Juncai Li, Ru Li, Yuxiang Zhou, Boxiang Ma, Jeff Z. Pan",
      "first_public": "2026-01-29",
      "added_at": "2026-10-18",
      "venue": "arXiv",
      "arxiv": "2601.21576",
      "paper_url": "https://arxiv.org/abs/2601.21576",
      "lens": [
        "analysis",
        "internalized",
        "learned-tokens"
      ],
      "representation": [
        "hidden-state"
      ],
      "computation": [
        "depth-wise"
      ],
      "training": [
        "supervised",
        "distillation"
      ],
      "questions": [
        "theory",
        "evaluation",
        "limitations",
        "efficiency"
      ],
      "evidence": [
        "formal",
        "empirical",
        "boundary"
      ],
      "modality": [
        "language",
        "symbolic"
      ],
      "summary": "Analyzes how skipping intermediate steps weakens learning signals in a simplified parity-task model. Introduces NatBool-DAG to test logical dependencies and ALiCoT, which supervises learned token states with intermediate logical labels or explicit-CoT representations.",
      "provenance": "Awesome-Latent-CoT; primary sources checked 2026-10-18",
      "author_names": [
        "Juncai Li",
        "Ru Li",
        "Yuxiang Zhou",
        "Boxiang Ma",
        "Jeff Z. Pan"
      ],
      "record_url": "https://latent-implicit-thinking.github.io/papers/2601.21576/",
      "legacy_record_url": "https://latent-implicit-thinking.github.io/papers/#paper-2601-21576",
      "citation_date": "2026-01-29",
      "json_url": "https://latent-implicit-thinking.github.io/papers/2601.21576/record.json",
      "markdown_url": "https://latent-implicit-thinking.github.io/papers/2601.21576/index.md",
      "bibtex_url": "https://latent-implicit-thinking.github.io/papers/2601.21576/citation.bib",
      "csl_url": "https://latent-implicit-thinking.github.io/papers/2601.21576/citation.csl.json"
    },
    {
      "id": "2602.08332",
      "title": "Latent Reasoning with Supervised Thinking States",
      "authors": "Ido Amos, Avi Caciularu, Mor Geva, Amir Globerson, Jonathan Herzig, Lior Shani, Idan Szpektor",
      "first_public": "2026-02-09",
      "venue": "arXiv",
      "arxiv": "2602.08332",
      "paper_url": "https://arxiv.org/abs/2602.08332",
      "lens": [
        "learned-tokens",
        "continuous"
      ],
      "representation": [
        "text-latent-hybrid",
        "continuous-state"
      ],
      "computation": [
        "token-wise",
        "recurrent"
      ],
      "training": [
        "supervised"
      ],
      "questions": [
        "capability",
        "efficiency"
      ],
      "evidence": [
        "empirical"
      ],
      "modality": [
        "language"
      ],
      "summary": "Supervises intermediate thinking states that bridge textual reasoning traces and latent computation.",
      "provenance": "Primary-source literature search, 2026-09-04",
      "author_names": [
        "Ido Amos",
        "Avi Caciularu",
        "Mor Geva",
        "Amir Globerson",
        "Jonathan Herzig",
        "Lior Shani",
        "Idan Szpektor"
      ],
      "record_url": "https://latent-implicit-thinking.github.io/papers/2602.08332/",
      "legacy_record_url": "https://latent-implicit-thinking.github.io/papers/#paper-2602-08332",
      "citation_date": "2026-02-09",
      "json_url": "https://latent-implicit-thinking.github.io/papers/2602.08332/record.json",
      "markdown_url": "https://latent-implicit-thinking.github.io/papers/2602.08332/index.md",
      "bibtex_url": "https://latent-implicit-thinking.github.io/papers/2602.08332/citation.bib",
      "csl_url": "https://latent-implicit-thinking.github.io/papers/2602.08332/citation.csl.json"
    },
    {
      "id": "2602.11683",
      "title": "ThinkRouter: Efficient Reasoning via Routing Thinking between Latent and Discrete Spaces",
      "authors": "Xin Xu, Tong Yu, Xiang Chen, Haoliang Wang, Julian McAuley, Saayan Mitra",
      "first_public": "2026-02-12",
      "venue": "arXiv, work in progress",
      "arxiv": "2602.11683",
      "paper_url": "https://arxiv.org/abs/2602.11683",
      "lens": [
        "continuous",
        "learned-tokens"
      ],
      "representation": [
        "text-latent-hybrid",
        "continuous-state"
      ],
      "computation": [
        "token-wise"
      ],
      "training": [
        "inference-only"
      ],
      "questions": [
        "capability",
        "efficiency",
        "controllability"
      ],
      "evidence": [
        "empirical"
      ],
      "modality": [
        "language"
      ],
      "summary": "Routes intermediate computation between latent and discrete representations to trade off reasoning quality, readability, and token cost.",
      "provenance": "Primary-source literature search, 2026-09-04",
      "author_names": [
        "Xin Xu",
        "Tong Yu",
        "Xiang Chen",
        "Haoliang Wang",
        "Julian McAuley",
        "Saayan Mitra"
      ],
      "record_url": "https://latent-implicit-thinking.github.io/papers/2602.11683/",
      "legacy_record_url": "https://latent-implicit-thinking.github.io/papers/#paper-2602-11683",
      "citation_date": "2026-02-12",
      "json_url": "https://latent-implicit-thinking.github.io/papers/2602.11683/record.json",
      "markdown_url": "https://latent-implicit-thinking.github.io/papers/2602.11683/index.md",
      "bibtex_url": "https://latent-implicit-thinking.github.io/papers/2602.11683/citation.bib",
      "csl_url": "https://latent-implicit-thinking.github.io/papers/2602.11683/citation.csl.json"
    },
    {
      "id": "2609.03807",
      "title": "Almost Free State Prediction Separation",
      "previous_titles": [
        "Free Pause Tokens"
      ],
      "authors": "John Langford, Nathan Godey, Giovanni Monea, Yoav Artzi, Harry Dong, Ying Fan, Gustavo de Rosa, Zheng Zhan",
      "first_public": "2026-09-03",
      "venue": "arXiv",
      "arxiv": "2609.03807",
      "paper_url": "https://arxiv.org/abs/2609.03807",
      "lens": [
        "learned-tokens",
        "beyond-sequential"
      ],
      "representation": [
        "hidden-state"
      ],
      "computation": [
        "parallel",
        "depth-wise"
      ],
      "training": [
        "pretraining"
      ],
      "questions": [
        "architecture",
        "capability",
        "efficiency"
      ],
      "evidence": [
        "empirical"
      ],
      "modality": [
        "language"
      ],
      "summary": "Moves pause-token computation into a parallel, weight-shared prediction stream, providing additional internal compute without increasing generated sequence length or key/value-cache length.",
      "provenance": "Primary-source literature search, 2026-09-05",
      "author_names": [
        "John Langford",
        "Nathan Godey",
        "Giovanni Monea",
        "Yoav Artzi",
        "Harry Dong",
        "Ying Fan",
        "Gustavo de Rosa",
        "Zheng Zhan"
      ],
      "record_url": "https://latent-implicit-thinking.github.io/papers/2609.03807/",
      "legacy_record_url": "https://latent-implicit-thinking.github.io/papers/#paper-2609-03807",
      "citation_date": "2026-09-03",
      "json_url": "https://latent-implicit-thinking.github.io/papers/2609.03807/record.json",
      "markdown_url": "https://latent-implicit-thinking.github.io/papers/2609.03807/index.md",
      "bibtex_url": "https://latent-implicit-thinking.github.io/papers/2609.03807/citation.bib",
      "csl_url": "https://latent-implicit-thinking.github.io/papers/2609.03807/citation.csl.json"
    },
    {
      "id": "2609.04489",
      "title": "Towards Understanding Pause Token Fine-Tuning Dynamics: A Mode Retention Perspective",
      "authors": "Jaehyeon Kim, Suhwan Kim, Nakyung Lee, Yeongoon Kim, Jimin Seo, Giho Lee, Jungwoo Lee",
      "first_public": "2026-09-03",
      "added_at": "2026-10-18",
      "venue": "arXiv",
      "arxiv": "2609.04489",
      "paper_url": "https://arxiv.org/abs/2609.04489",
      "summary": "Places pause tokens at reasoning-step boundaries and masks their training loss. Studies how this affects learning new tasks while retaining earlier capabilities, including during reinforcement learning.",
      "lens": [
        "learned-tokens",
        "analysis"
      ],
      "representation": [
        "discrete-token",
        "text-latent-hybrid"
      ],
      "computation": [
        "token-wise"
      ],
      "training": [
        "supervised",
        "reinforcement-learning"
      ],
      "questions": [
        "capability",
        "generalization",
        "interpretability"
      ],
      "evidence": [
        "empirical",
        "probing"
      ],
      "modality": [
        "language",
        "symbolic"
      ],
      "provenance": "Primary-source literature search, 2026-10-18",
      "author_names": [
        "Jaehyeon Kim",
        "Suhwan Kim",
        "Nakyung Lee",
        "Yeongoon Kim",
        "Jimin Seo",
        "Giho Lee",
        "Jungwoo Lee"
      ],
      "record_url": "https://latent-implicit-thinking.github.io/papers/2609.04489/",
      "legacy_record_url": "https://latent-implicit-thinking.github.io/papers/#paper-2609-04489",
      "citation_date": "2026-09-03",
      "json_url": "https://latent-implicit-thinking.github.io/papers/2609.04489/record.json",
      "markdown_url": "https://latent-implicit-thinking.github.io/papers/2609.04489/index.md",
      "bibtex_url": "https://latent-implicit-thinking.github.io/papers/2609.04489/citation.bib",
      "csl_url": "https://latent-implicit-thinking.github.io/papers/2609.04489/citation.csl.json"
    },
    {
      "id": "2609.21605",
      "title": "Trading Depth for Time in Recurrent Transformers",
      "authors": "Zeyi Huang, Xuehai He, Yelong Shen, Yong Jae Lee",
      "first_public": "2026-09-18",
      "added_at": "2026-10-18",
      "venue": "arXiv",
      "arxiv": "2609.21605",
      "paper_url": "https://arxiv.org/abs/2609.21605",
      "lens": [
        "recurrent",
        "learned-tokens"
      ],
      "representation": [
        "hidden-state",
        "continuous-state"
      ],
      "computation": [
        "recurrent",
        "token-wise",
        "stateful"
      ],
      "training": [
        "pretraining"
      ],
      "questions": [
        "architecture",
        "efficiency",
        "scaling"
      ],
      "evidence": [
        "empirical",
        "scaling"
      ],
      "modality": [
        "language"
      ],
      "summary": "Inserts a hidden-state refinement step between consecutive text tokens in a recurrent transformer. Compares this extra thinking time with adding physical layers or transformer loops under matched computation budgets.",
      "provenance": "Primary-source literature search, 2026-10-18",
      "author_names": [
        "Zeyi Huang",
        "Xuehai He",
        "Yelong Shen",
        "Yong Jae Lee"
      ],
      "record_url": "https://latent-implicit-thinking.github.io/papers/2609.21605/",
      "legacy_record_url": "https://latent-implicit-thinking.github.io/papers/#paper-2609-21605",
      "citation_date": "2026-09-18",
      "json_url": "https://latent-implicit-thinking.github.io/papers/2609.21605/record.json",
      "markdown_url": "https://latent-implicit-thinking.github.io/papers/2609.21605/index.md",
      "bibtex_url": "https://latent-implicit-thinking.github.io/papers/2609.21605/citation.bib",
      "csl_url": "https://latent-implicit-thinking.github.io/papers/2609.21605/citation.csl.json"
    },
    {
      "id": "2609.30783",
      "title": "Skip the Talk, Re-Focus on Vision: Latent Reasoning for Reasoning Segmentation in Multimodal Large Language Models",
      "authors": "Tianhang Guo, Yulin He, Wei Chen, Wenjuan Zhou, Yuhang Li, Xinbiao Gan",
      "first_public": "2026-09-25",
      "venue": "arXiv",
      "arxiv": "2609.30783",
      "paper_url": "https://arxiv.org/abs/2609.30783",
      "lens": [
        "learned-tokens"
      ],
      "representation": [
        "discrete-token",
        "hidden-state",
        "multimodal-latent"
      ],
      "computation": [
        "token-wise"
      ],
      "training": [
        "supervised",
        "reinforcement-learning"
      ],
      "questions": [
        "capability",
        "efficiency"
      ],
      "evidence": [
        "empirical"
      ],
      "modality": [
        "multimodal",
        "vision"
      ],
      "summary": "LIRSeg generates a fixed sequence of learned special tokens before a target description and bounding box used for segmentation. Spatial alignment grounds the tokens' hidden states in relevant image regions, followed by training with segmentation rewards.",
      "provenance": "Primary-source literature search, 2026-10-18",
      "added_at": "2026-10-18",
      "author_names": [
        "Tianhang Guo",
        "Yulin He",
        "Wei Chen",
        "Wenjuan Zhou",
        "Yuhang Li",
        "Xinbiao Gan"
      ],
      "record_url": "https://latent-implicit-thinking.github.io/papers/2609.30783/",
      "legacy_record_url": "https://latent-implicit-thinking.github.io/papers/#paper-2609-30783",
      "citation_date": "2026-09-25",
      "json_url": "https://latent-implicit-thinking.github.io/papers/2609.30783/record.json",
      "markdown_url": "https://latent-implicit-thinking.github.io/papers/2609.30783/index.md",
      "bibtex_url": "https://latent-implicit-thinking.github.io/papers/2609.30783/citation.bib",
      "csl_url": "https://latent-implicit-thinking.github.io/papers/2609.30783/citation.csl.json"
    },
    {
      "id": "2609.33134",
      "title": "Ceiling of a Task: When Can a Transformer Succeed Without Its Chain of Thought?",
      "authors": "Jiashu He, Jinxuan Fan, Xiao Xiao, Radu Marculescu, Alejandro Ribeiro",
      "first_public": "2026-09-27",
      "added_at": "2026-10-18",
      "venue": "arXiv",
      "arxiv": "2609.33134",
      "paper_url": "https://arxiv.org/abs/2609.33134",
      "lens": [
        "analysis",
        "learned-tokens"
      ],
      "representation": [
        "discrete-token"
      ],
      "computation": [
        "token-wise"
      ],
      "training": [
        "analysis-only"
      ],
      "questions": [
        "theory",
        "limitations",
        "evaluation"
      ],
      "evidence": [
        "formal",
        "empirical",
        "boundary"
      ],
      "modality": [
        "language",
        "symbolic"
      ],
      "summary": "Analyzes tasks for which a fixed-depth transformer needs an informative chain of thought to exceed a task-specific accuracy limit. Combines circuit-based results with experiments that replace, erase, or shuffle reasoning tokens, including comparisons with filler tokens.",
      "provenance": "Primary-source literature search, 2026-10-18",
      "author_names": [
        "Jiashu He",
        "Jinxuan Fan",
        "Xiao Xiao",
        "Radu Marculescu",
        "Alejandro Ribeiro"
      ],
      "record_url": "https://latent-implicit-thinking.github.io/papers/2609.33134/",
      "legacy_record_url": "https://latent-implicit-thinking.github.io/papers/#paper-2609-33134",
      "citation_date": "2026-09-27",
      "json_url": "https://latent-implicit-thinking.github.io/papers/2609.33134/record.json",
      "markdown_url": "https://latent-implicit-thinking.github.io/papers/2609.33134/index.md",
      "bibtex_url": "https://latent-implicit-thinking.github.io/papers/2609.33134/citation.bib",
      "csl_url": "https://latent-implicit-thinking.github.io/papers/2609.33134/citation.csl.json"
    },
    {
      "id": "2610.20641",
      "title": "AbridgeKV: Compressing Long Reasoning into the KV Cache of Learned Pause Tokens",
      "authors": "Jack Cai, Lijie Yang, Vasilis Kontonis, Dimitris Papailiopoulos, Shivam Garg",
      "first_public": "2026-10-07",
      "added_at": "2026-10-18",
      "venue": "arXiv",
      "arxiv": "2610.20641",
      "paper_url": "https://arxiv.org/abs/2610.20641",
      "summary": "Compresses each completed block of generated reasoning into the key/value cache of learned pause tokens, then discards the block's original cache. Later tokens attend to the compact replacement.",
      "lens": [
        "learned-tokens"
      ],
      "representation": [
        "cache-memory",
        "text-latent-hybrid"
      ],
      "computation": [
        "token-wise",
        "stateful"
      ],
      "training": [
        "supervised"
      ],
      "questions": [
        "efficiency",
        "architecture"
      ],
      "evidence": [
        "empirical"
      ],
      "modality": [
        "language",
        "symbolic"
      ],
      "provenance": "Primary-source literature search, 2026-10-18",
      "author_names": [
        "Jack Cai",
        "Lijie Yang",
        "Vasilis Kontonis",
        "Dimitris Papailiopoulos",
        "Shivam Garg"
      ],
      "record_url": "https://latent-implicit-thinking.github.io/papers/2610.20641/",
      "legacy_record_url": "https://latent-implicit-thinking.github.io/papers/#paper-2610-20641",
      "citation_date": "2026-10-07",
      "json_url": "https://latent-implicit-thinking.github.io/papers/2610.20641/record.json",
      "markdown_url": "https://latent-implicit-thinking.github.io/papers/2610.20641/index.md",
      "bibtex_url": "https://latent-implicit-thinking.github.io/papers/2610.20641/citation.bib",
      "csl_url": "https://latent-implicit-thinking.github.io/papers/2610.20641/citation.csl.json"
    }
  ]
}