Hide papers I cannot work with
Change how "my score" is weighted — equal weights by default; there is no built-in ranking
Suggestions
Share, export, reset
What the numbers and labels mean, and where they come from
No code was run to produce anything on this page. Scripts read each paper's PDF and a shallow copy of its repository. Nobody has measured how well these scores predict whether a paper reproduces. Click a paper to see the evidence behind its numbers.
| Label | Who assigned it | How it reached this page |
|---|---|---|
| Oral / spotlight / poster, track | The NeurIPS 2026 programme committee | Copied unchanged from the hackathon's official claim list |
| Standard / "harder than most", the claim to verify, table-or-figure, metric | The hackathon organisers | Copied unchanged from the official claim list (hosted on Hugging Face) |
| Topics and areas (dashed, italic) | Nobody — computed by this tool from the text | Automatic and rough. A paper can carry several, or none |
| Turnkey, checkable, read time, match, my score | Nobody — heuristics of this tool | Rules are listed in the project README |
Nothing here comes from the paper authors' own keywords (the claim list has none) or from Papers with Code. That site has its own categories for NeurIPS 2026 papers; they are not used here because the site does not permit automated collection. Only the link to each paper's page is shown.
Turnkey 0–10: how runnable the repository looks. Starts at 3; +1 per good sign (Hugging Face data or weights loaded in code, lock file, Dockerfile, eval script, reproduce script); −1 or −2 per friction (Google Drive downloads, C++/CUDA build, gated data, video scraping, no dependency file, very short README). 0 only when there are no code files, the README says the code is not released, or the method under test is a hosted closed model. "no repo found" = no GitHub link in the PDF or on its project page; the code may still exist.
Checkable 0–10: how cleanly you can tell whether you matched the claim (exact numbers in a table, a reported spread, few benchmarks). Read: estimated minutes for abstract + method + setup.
Match: an exact match means your word appears (at the start of a word) in the title (100%) or in the claim, method or background (70%). A related match (up to 60%) means the paper's text is unusually close in meaning to your word, judged by a text-embedding model; it is approximate and can be wrong.
Similarity (the "sim" numbers and the lines on the map) is the cosine similarity between text embeddings of two papers' title + claim. In this collection two random papers score about 0.68 and a paper's nearest neighbours 0.80–0.90. High similarity means similar wording and subject. It does not mean shared code, data or authors.