About Text Diffusions

How to use it

1. Build a practical shortlist

   Start with core papers on inference acceleration that have a code link. A link does not establish that the code reproduces the paper or fits your hardware.

   Find acceleration papers with code

2. Find a known paper or a phrase

   Search by its original title or arXiv ID. Quotation marks require a phrase match; this example searches for “masked language”. A valid search can return no results.

   Try a quoted search

3. Compare before committing time

   Select Compare on up to four paper cards. Check the dataset, metric and protocol, then follow the original sources. A comparison URL preserves the selection. Paper pages also offer BibTeX, RIS and Markdown exports.

   Choose papers to compare

4. Check what changed

   Open a dated weekly summary to see recorded additions and source or code changes. These are observed changes, not a claim that every important paper has been collected.

   Read weekly changes

5. Follow one topic

   Add the inference-acceleration RSS feed to your reader. RSS does not require an account or email signup.

   Open the topic RSS feed

6. Learn the direction

   Choose a starting route: foundations, inference acceleration or controllability. Follow its reading steps and use the glossary when a term is unfamiliar.

   Choose a learning path

Sources and selection

Each paper has an original-source link. Titles and abstracts retain their original wording and language; interface text and available explanations follow the selected language. Coverage depends on the collected sources and is not exhaustive.

Initial scope and topic labels use the title and abstract. Ambiguous cases can receive an additional model-assisted classification when processing is available. These labels are a navigation aid, not a measured guarantee of relevance.

core: directly relevant diffusion language model research. transferable: potentially reusable ideas from adjacent tasks. out_of_scope: outside the focus. uncertain: insufficient or ambiguous evidence. The home selection uses core papers; the paper search exposes all four scopes.

The eight starting topics are masked/discrete diffusion, latent diffusion, inference acceleration, long context, reasoning, post-training, controllability and code generation. Topic membership is not an endorsement or a ranking of scientific quality.

What an automated review means

An automated review is a reading aid, not peer review. The validator checks that quoted evidence occurs in the available source text and is bound to its source URL and hash. A source-matched quotation does not prove that a claim is scientifically correct.

Extracted facts and model interpretations are distinguished. Reported numerical results need a value, dataset, metric, protocol and supporting excerpt. The site does not run its own experiments or independently reproduce the reported numbers.

A card may have an abstract alone, missing full text or no accepted analysis. Missing information stays unknown; an unsuccessful analysis does not remove the original-source card. Do not infer a limitation has been ruled out merely because it is not listed.

Comparison tables warn when dataset, metric or protocol information is missing or differs. Even matching labels do not establish equivalent hardware, data splits, compute budgets or statistical confidence: check the original papers.

No new model call is triggered by reading, searching, comparing or asking the corpus. Corpus answers retrieve stored analyses and source excerpts; they are not an unrestricted research assistant and can lack evidence for your question.

Dates and updates

Publication date, successful collection time and snapshot creation time mean different things. A new snapshot is not proof of a new publication. Weekly pages describe the seven UTC dates ending on their stated date and changes observed in the stored history.

The home page shows zero to five recent core papers. When none qualify, it can show a separately dated historical selection. If processing is unavailable, existing published content remains readable; completeness and freshness are not guaranteed.

Access and personal choices

Public reading, search, comparisons and citation exports do not require registration. Without an account, saved papers and topic choices stay in this browser; account features provide synchronization when available. Browser-local choices are not a backup.

RSS is available separately from email and Telegram. Delivery channels remain off until configured; check the subscription page for current availability. An email subscription requires confirmation and can be cancelled.

If a summary or label looks wrong, use the linked original source before relying on it. The site helps you navigate the literature; it does not replace your assessment of methods, assumptions and results.