Methodology

Paper of the day

Every day just after midnight UTC, Claude reads a shortlist of about 45 papers: the top 30 of the to reproduce ranking and the top 15 hidden gems from the last two weeks. For each one it sees the abstract, authors, community activity, code and reproduction status, and the compute estimate. It also sees the last ten picks. It chooses the one paper that best fits the criteria below, and writes a short note on why it's worth reproducing and what you'd learn.

1. 1.Teaches a transferable technique. The core idea is a reusable building block - a loss, a training recipe, an architectural change, a data method or an inference trick - that you could carry into other work, not a one-off system.
2. 2.Reproducible on modest hardware. The central claim can be tested on a laptop or a single GPU with public models and data, even if only at reduced scale.
3. 3.A clear, checkable claim. There is a specific result to match - a number, a curve, an ablation - so you can tell whether your reproduction worked.
4. 4.Builds fundamentals. Reproducing it exercises core skills (training loops, attention internals, RL objectives, careful evaluation) and sharpens intuition for why methods work, so you can later try your own variations.
5. 5.Interesting and timely. The idea is surprising, elegant or likely to matter, and people are paying attention to it or soon will be.
6. 6.Specified well enough to implement. The method is described precisely enough to build. Official code helps as a reference to check against, but isn't required.
7. 7.Something different. Recent picks are taken into account so the daily paper rotates across areas instead of repeating one topic.

A paper can only be paper of the day once, ever. The pick is stored when it's made, so it stays the same all day. Popularity is treated as context rather than the deciding factor, so a quieter paper that fits better can win. If Claude is unavailable, the day's paper falls back to the top reproduce pick that runs on a single GPU or less; the label on the home page says which method was used.

The lists

To reproduce

60% discussion, 25% compute feasibility and 15% paper type. Discussion blends HF Daily Papers upvotes, Hacker News points and comments, official-repo stars and citations, each log-scaled against the top paper in the timeframe. Feasibility is the compute needed for the smallest run that tests the paper's claim, in GPU-hours on one A100-class GPU: laptop up to 2, single GPU up to a week. For the top papers Claude reads the hardware and training budget from the full text; for the rest it's estimated from model sizes and token counts in the abstract. Reproductions only count GitHub repos about this paper, not code that just cites it. Surveys and position papers are left out.

Hidden gems

Papers that haven't taken off yet but show signs they could. The signals are a concept suddenly appearing in several new papers, fast early upvotes, authors or a lab with a recent well-received paper, and being a runnable method with code. The more attention a paper already has, the more its score is discounted, and the 10% most-discussed papers are left out.

To read

Discussion alone.

Votes and comments on unarXived add a small boost of at most +0.1 that applies immediately. Source data is refreshed every three hours, with a full re-ingest daily.