Skip to main content

Download the reading list

Get citations for your reference manager, a text copy to read, or JSON for research tools and agents. Every download uses the same paper list as the website.

Using the JSON files in a script or agent

Start with papers.json. It contains papers, precursors (the background papers), and metadata. The agent guide links to every file.

Paper details
id is a stable identifier. author_names preserves author order. paper_url links to the source; record_url links to an individual page with its own JSON, Markdown, BibTeX and CSL-JSON downloads. legacy_record_url preserves the earlier full-list anchor.
Published versions
Optional publication metadata gives a separately verified conference or journal version, with its own title, authors, year and citation links. Missing metadata means we have not verified that version, not that it does not exist.
Demos
Optional demos link to interactive examples and model details. relationship distinguishes an implementation from a related_method example; paper_relationships names the associated papers. A related-method demo is not an implementation of the current paper. Demo outputs are not benchmark evidence, and demos do not replace paper citations.
Dates
first_public is the earliest dated public version we found, usually the first arXiv submission. citation_date is the initial arXiv submission date used in citation exports. Earlier conference presentations have a separate source note.
Following additions
added_at is when a paper joined this list, not when it was published. It is tracked for additions after the initial collection; initial records have no added_at. Filter this field to find new entries, including older papers we previously missed. Release history also lists added_papers by stable ID.
Keeping a local copy current
Release history can identify updated_papers with changed fields and removed_papers by ID. The current JSON file is the authority for which entries are on the list.
Categories and summaries
lens lists approaches. Other fields describe training, computation, research questions, evidence, and modality. A paper can have several labels. Summaries and labels are written for this list; they are not the authors' abstracts.
Sources
provenance records where we found the paper. Use paper_url to check the paper itself.

Category definitions (ontology.json) · Field types and allowed values (JSON Schema) · Version history (JSON)

Download one topic

To follow updates, see recent additions, subscribe to the RSS feed, or read what changed.

Citing and reusing the list

Cite the original papers for research claims. The complete papers.bib and papers.csl.json downloads cite arXiv preprints. A separate BibTeX file and CSL-JSON file contain the subset with verified published-version records. Individual paper pages offer both choices where available. Check the source and cite the version you used.

If you reuse our categories, summaries, or data, credit Implicit Chain-of-Thought, version 0.9.4 (2026), https://latent-implicit-thinking.github.io/, with your access date. Our original text and annotations are available under CC BY 4.0. Papers, code, and other third-party material keep their own licenses.

Date and citation-format details

Dates usually refer to the first arXiv submission, not the public-announcement time or conference year. Where an earlier presentation is documented, its date is shown separately from the linked arXiv submission date.

Source links are not pinned to an arXiv revision. Author names are kept in full and in order. CSL-JSON stores each name as a literal string rather than guessing which parts are given and family names.

Thanks to the Awesome-Latent-CoT bibliography and its associated survey, Reasoning Beyond Language, which helped us find many of these papers.

Suggest a paper or report a correction →