# Implicit Chain-of-Thought > A reading list about reasoning without writing every intermediate step, with explanations, paper summaries, and downloadable citations. This research is also discussed as latent reasoning or latent chain-of-thought. Yuntian Deng and colleagues first proposed implicit chain-of-thought reasoning in [2023](https://arxiv.org/html/2311.01460v1#S2.SS3): training language models to reason through continuous hidden states rather than write out explicit intermediate steps. The paper's [distillation method](https://arxiv.org/html/2311.01460v1#S3) is one implementation of that proposal. Their [2024 work](https://arxiv.org/abs/2405.14838) introduced step-by-step internalization. Version: 0.9.4 Website updated: 2026-10-18 Literature checked through: 2026-10-18 Papers: 207 Related background papers: 8 ## Read and search - [Overview](https://latent-implicit-thinking.github.io/) - [Start here](https://latent-implicit-thinking.github.io/topics/): explanations and examples - [Browse papers](https://latent-implicit-thinking.github.io/papers/): search, filters, and date order - [Recent additions](https://latent-implicit-thinking.github.io/recent/): papers added after the initial collection, grouped by the date they joined this list - [Complete Markdown reading list](https://latent-implicit-thinking.github.io/data/reading-list.md): all papers, authors, summaries, and source links - [Downloads](https://latent-implicit-thinking.github.io/data/): citation files and JSON documentation ## Topics - [Internalized reasoning](https://latent-implicit-thinking.github.io/topics/internalized-reasoning/): Learn from worked solutions so that some or all of the reasoning can later happen without writing those steps. Training may use a separate teacher model or progressively remove written steps from the model being trained. [Markdown](https://latent-implicit-thinking.github.io/data/topics/internalized-reasoning.md) · [JSON](https://latent-implicit-thinking.github.io/data/topics/internalized-reasoning.json) - [Thinking tokens](https://latent-implicit-thinking.github.io/topics/thinking-tokens/): Give a model extra tokens for computation, such as pauses, filler symbols, or learned planning tokens. These tokens need not spell out the solution. [Markdown](https://latent-implicit-thinking.github.io/data/topics/thinking-tokens.md) · [JSON](https://latent-implicit-thinking.github.io/data/topics/thinking-tokens.json) - [Continuous-state reasoning](https://latent-implicit-thinking.github.io/topics/continuous-states/): Use vectors—lists of numbers—to carry intermediate reasoning from one step to the next, instead of selecting a text token at every step. Some methods use the model's hidden states; others combine several candidate token representations. [Markdown](https://latent-implicit-thinking.github.io/data/topics/continuous-states.md) · [JSON](https://latent-implicit-thinking.github.io/data/topics/continuous-states.json) - [Recurrent reasoning](https://latent-implicit-thinking.github.io/topics/recurrent-reasoning/): Repeat parts of the model or keep intermediate information in memory while working through a problem. Some methods also learn how many passes to run. [Markdown](https://latent-implicit-thinking.github.io/data/topics/recurrent-reasoning.md) · [JSON](https://latent-implicit-thinking.github.io/data/topics/recurrent-reasoning.json) - [Alternative generation methods](https://latent-implicit-thinking.github.io/topics/alternative-reasoning-schedules/): Change how intermediate steps are generated, revised, or searched—for example, by refining several steps in parallel. These are related approaches: some still produce written chains of thought. [Markdown](https://latent-implicit-thinking.github.io/data/topics/alternative-reasoning-schedules.md) · [JSON](https://latent-implicit-thinking.github.io/data/topics/alternative-reasoning-schedules.json) - [Analysis and evaluation](https://latent-implicit-thinking.github.io/topics/implicit-reasoning-evaluation/): Measure how well these methods work, investigate what their internal computation does, and identify where they fail. This includes experiments, studies of model internals, and mathematical results. [Markdown](https://latent-implicit-thinking.github.io/data/topics/implicit-reasoning-evaluation.md) · [JSON](https://latent-implicit-thinking.github.io/data/topics/implicit-reasoning-evaluation.json) ## Files for scripts and research agents - [JSON paper list](https://latent-implicit-thinking.github.io/data/papers.json): metadata, papers, and a separate background collection - [Category definitions](https://latent-implicit-thinking.github.io/data/ontology.json): category IDs, descriptions, and allowed filter values - [JSON Schema](https://latent-implicit-thinking.github.io/data/schema.json): field types and constraints - [BibTeX](https://latent-implicit-thinking.github.io/data/papers.bib) and [CSL-JSON](https://latent-implicit-thinking.github.io/data/papers.csl.json): citations for all papers, including background - [Published-version BibTeX](https://latent-implicit-thinking.github.io/data/publications.bib) and [CSL-JSON](https://latent-implicit-thinking.github.io/data/publications.csl.json): only the subset with verified publication records; not a replacement for the complete preprint files - [Changes](https://latent-implicit-thinking.github.io/data/changes.json) and [RSS feed](https://latent-implicit-thinking.github.io/feed.xml): website and reading-list updates - [Source repository](https://github.com/latent-implicit-thinking/latent-implicit-thinking.github.io) ## Using the data Use the original papers for scientific claims. Summaries and categories are written for this reading list, not copied abstracts. Papers can belong to several categories. Coverage includes related generation methods and is not exhaustive. - `id` identifies a paper; `paper_url` links to the original source; `record_url` links to an independently fetchable entry at `/papers/{id}/`. Earlier full-list anchors still work through `legacy_record_url`. - Each entry has `json_url`, `markdown_url`, `bibtex_url`, and `csl_url`. Use the topic downloads or individual entries when you do not need the complete list. - `author_names` preserves the authors' names and order. - Optional `related_terms` provides editorial search terminology, such as latent reasoning or latent chain-of-thought. These terms are also visible on the paper's page; they do not replace its title or citation. Search them alongside the title and summary. - Optional `method_comparisons` contains the comparisons shown on a paper's page, with the papers discussed in `source_ids`, primary-source links in `source_urls`, and readable `source_labels`. Each comparison identifies its topic guide. A comparison may explain a difference or a documented influence; its presence is not itself an influence claim. - `added_at` is when a paper was added to this list, not its publication date. Tracking began 2026-10-18; initial-collection records omit this field. Release history lists `added_papers` IDs. A new addition may be an older paper. - `first_public` is the earliest public date we have verified, usually the first arXiv submission. An earlier conference presentation has a source note. It is not an independently verified announcement date. - `citation_date` is the initial submission date of the linked arXiv preprint. BibTeX and CSL-JSON describe that preprint, not a later proceedings version. Source URLs may serve revised versions. - Optional `publication` metadata describes a separately verified published version. Its own title, author list and year may differ; use `publication_bibtex_url` or `publication_csl_url` for that version. Absence means we have not verified a publication record, not that the paper is unpublished. - Optional `demos` link to interactive examples and model details. Each demo's `paper_relationships` distinguishes `implementation` from `related_method`; `relationship` gives its connection to the current paper. A related-method demo does not implement that paper. Topic downloads also list the demos shown in the guide. Demo outputs are examples, not benchmark evidence; use the method paper for scientific claims and cite the model separately when you use it. - Release entries can list `updated_papers` (IDs and changed fields) and `removed_papers` (IDs). Refresh cached records using the current corpus; do not treat a historical release as the current membership list. - `provenance` says where we first found a paper. - Category IDs and filter values are defined in the category file; related terms are not necessarily synonyms. ## Research workflow Choose the relevant topic, identify papers whose methods address the question, then read the original sources. The topic guides compare training and computation choices with source links. Verify the exact result, setting and limitations before citing it. Chronological order is not proof of influence, and citing a paper does not necessarily mean building on its method. Export a citation for the version actually used. This curated list is a starting point, not an exhaustive literature search. ## Corrections and credit - [Suggest a paper or correction](https://latent-implicit-thinking.github.io/contribute/) - [Changelog](https://latent-implicit-thinking.github.io/changelog/) - [Citation and reuse guidance](https://latent-implicit-thinking.github.io/data/#citation) - [CC BY 4.0](https://latent-implicit-thinking.github.io/license/): applies to our categories, summaries, and data; original papers and code keep their own terms ## Acknowledgment Thanks to the [Awesome-Latent-CoT bibliography](https://github.com/EIT-NLP/Awesome-Latent-CoT) and its associated survey, which helped us find many of these papers.