arXiv's cs.AI feed alone runs to hundreds of submissions a day. Reading it is not a plan.

Three filters, in order

Attention first. Community-ranked lists — Hugging Face Daily Papers, researchers who post consistently — are a cheap prior. Not authoritative, but a better starting distribution than chronological.

Problem second. Keyword search finds papers sharing your words; semantic search finds papers sharing your problem. For a literature review the second is what you want, and the two disagree more than people expect.

Structure third. Read the abstract, the limitations, and one figure. If it survives, it earns a note with a quote and a page number — not a summary. Summaries decay; quotes stay checkable.

What to skip deliberately

Anything you cannot state a reason for reading. "It seemed relevant" is how a triage system turns back into a pile.

The honest limit

No workflow makes you current with a field producing hundreds of papers daily. The goal is a defensible sample, not coverage.

Filter one: attention as a prior

Community-ranked lists are not authoritative, and treating them as such is a mistake. But as a prior over what to look at first, they beat chronological order substantially.

Hugging Face Daily Papers, a handful of researchers who post consistently, and subreddit activity all encode a rough judgement from people closer to the work than you are. Use them to order the queue, not to decide the queue.

Filter two: problem, not words

Keyword search retrieves papers that share your vocabulary. Semantic search retrieves papers that share your problem. These disagree constantly, and the disagreement is informative.

A paper solving your exact problem in a different subfield's language is invisible to keyword search and is often the most valuable thing you could read. Conversely, a paper sharing all your terminology may be addressing something unrelated.

Run both. Read the disagreements first.

Filter three: structured skim

Three parts, in this order. Abstract — what is claimed. Limitations — what the authors admit, which is where the honest signal is. One figure — usually the main result, which tells you whether the claim survives contact with data.

If it survives that, it earns a note. A note is a quote and a page number, not a summary. Summaries decay into your paraphrase of your paraphrase; quotes stay checkable against the source three months later.

What to skip on purpose

Anything you cannot state a reason for reading in one sentence. "Seemed relevant" is how a triage system silently reverts to a reading pile.

The honest limit

No workflow makes you current with a field producing hundreds of papers a day. Coverage is not achievable and pursuing it is how people burn out. The goal is a defensible sample — one you could justify to a reviewer — plus a mechanism for noticing when you have missed something that matters.

That mechanism is usually other people, not a tool.

The workflow as a diagram

The order is deliberate. Each filter is cheaper than the one after it, so the expensive step — actually reading — only runs on what survived three cheap ones.