The reading-load math for an active researcher in mid-2026: roughly 200 papers per day land on arXiv across the categories most ML researchers track. Filtered to one subfield it is 30-50 per week worth opening. Most researchers we know who keep up read 5-10 of those 30-50 per week.
The question is not "how do you read everything" because nobody does. The question is "how do you decide which 5-10 in 90 minutes". The workflow below is what we have seen survive past 6 months on the researchers who maintain it. It runs Sunday morning, takes 90 minutes total, and produces a ranked list of 1-2 papers to deep-read during the week plus 3-5 to skim plus a clean queue for next Sunday.
Borrowed from researchers who write us. Not invented.
Step 1 — Friday night: dump the week's papers into one queue (10 min)
Every paper-worth-opening that hit your feeds during the week goes into one inbox. Could be Zotero with a "Triage" collection, an Obsidian folder, a Notion database, or a plain text file. The format does not matter; the constraint matters.
The constraint: nothing gets deep-read on Friday. Nothing gets deep-read mid-week. The job Friday night is intake only. Decide nothing.
What goes in: papers you saw on X, papers from your arXiv RSS feeds, papers from editorial newsletters (TLDR AI, The Batch, Alpha Signal, DIGEST — whichever you read), papers from author whitelists, papers your advisor or PI sent. Anything where your future self might say "I should read this".
What does not go in: papers you already know you will not read (skip the dump entirely), papers you read mid-week and finished (those are done), papers you started and bounced on (move to "rejected" with a one-line note).
By Friday end-of-day the queue has 20-60 papers depending on your week. Sleep on it.
Step 2 — Saturday morning: priority-sort by 3 signals (15 min)
Saturday is the cool-down. The Friday excitement about "oh this looks great" has worn off. You can rank more honestly. Sort the queue using 3 signals only, in this order:
Signal 1 — Does this connect to active work? Your current project, your current paper, the literature review you owe. Papers that connect to active work get priority above papers that just look interesting. Mark these "active".
Signal 2 — Is the author someone whose work you have read and learned from? Track record matters. A new paper from someone whose past 3 papers you found useful is higher-priority than a new paper from a name you do not recognize. Mark these "trusted".
Signal 3 — Does the abstract suggest a result that would change something in your mental model? This is the surprise-signal. If the abstract says "we found that X is not necessary for Y" and you currently believe X is necessary for Y, that is high-priority regardless of author. Mark these "surprise".
Papers with no signal flagged go to the bottom. Papers with 2+ signals go to the top. Papers with 1 signal go to the middle, sorted by which signal (active > trusted > surprise).
Do not read deep yet. Just rank.
Step 3 — Sunday 9am: speedrun the abstracts (20 min)
For each paper in priority order, read the abstract only. Set a 30-second timer per abstract. Force the speed.
For each abstract, decide one of three things:
- Read further — the abstract earns the next step (figures + intro). About 10-15 papers should make it through.
- Skim only — the abstract gave you what you need; you do not need to open the paper. Add a one-line note to your reference manager and move on. About 15-25 papers.
- Reject — the abstract is off-topic or the result is uninteresting. Archive with a one-line note (Step 7) and stop. About 10-20 papers.
The 30-second timer is real. If you cannot decide in 30 seconds, default to "skim only" and move on. The papers that need more than 30 seconds to triage are the papers you should not be triaging — they are the papers you would have read anyway.
Step 4 — Sunday 9:30am: figure-only sweep on top 10 (15 min)
The 10-15 papers that made "read further" get a figure-only pass. Open each, scroll through the figures, read the figure captions. Do not read the prose body.
Figures tell you the contribution faster than prose. Figure 1 is usually the architecture or the result framing. Figures 3-5 are usually the experimental tables. The last figure is often the ablation that tells you what the authors actually believe matters.
For each paper, decide:
- Deep-read this week — the figures suggest a contribution worth the full read. 1-3 papers max.
- Intro-only further — figures look promising but unclear. Go to Step 5.
- Skim only — figures revealed it is incremental or outside your interest. Add the one-line note and move on.
The "deep-read this week" pile is the protected output of this whole workflow. The other 47 papers exist to surface those 1-3.
Step 5 — Sunday 9:45am: intro paragraphs on top 5 (15 min)
The papers that came out of Step 4 as "intro-only further" plus the "deep-read this week" picks get an intro-paragraph pass. Read the first paragraph of the introduction (after the abstract you already read), the last paragraph of the introduction, and the conclusion section's first paragraph.
This catches papers where the figures looked promising but the authors are honest about a major caveat (intro paragraph 1 often hedges), or papers where the figures looked incremental but the conclusion connects to something larger.
Refine the "deep-read this week" pile. Should still be 1-3 papers max. If you have 5+ you are not triaging hard enough.
Step 6 — Sunday 10am: pick 1-2 for deep read this week (5 min)
The decision moment. From your 1-3 candidates, commit to 1 (if you are time-constrained) or 2 (if you have reading time and the papers connect to each other or to your active work).
Write the deep-read papers into your calendar. Block 60-90 minutes for each. If you do not block time, the deep-read will not happen and the triage was wasted.
The papers you did not pick from the candidate pile go to "skim only" (with a one-line note about what you would have wanted to know from a deeper read).
Step 7 — Sunday 10:15am: archive the rejected with a one-line note (10 min)
The papers that hit "skim only" or "reject" get a one-line note in your reference manager before they exit your active workflow. The note is not a summary — it is a future-search-trigger.
Format: [paper-title-or-short-form] [skim/reject] [one-line why or what]
Examples:
Scaling laws for downstream tasks (Hernandez 2024) skim Confirms Chinchilla-style scaling holds for finetuning regime, no surpriseMixtral mid-training analysis (Marigold 2025) reject Wrong subfield (we are in interpretability, this is training infrastructure)Constitutional AI v3 (Anthropic 2026) skim Read intro only, the v3 changes are incremental to v2 for our purposes
These notes are how you find the paper 6 months from now when something connects. The notes are short by design. If you are writing more than one line per paper, you are reading more than you intended.
The workflow ends with the queue empty, the deep-read pile in your calendar, and the reference manager updated.
Why 90 minutes specifically
The 90 minutes is not arbitrary. It is the longest single block of focused reading-triage we have seen researchers maintain weekly past 6 months. 2-hour weekly blocks slip on busy weeks and then never recover. 60-minute blocks rush Steps 4-5 and the deep-read picks get worse.
90 minutes lets you give 60 seconds to most papers, 2-3 minutes to candidate papers, and 30 seconds for the rejected-with-note pass. It fits between two coffees on a Sunday morning. It does not eat your weekend.
What this workflow assumes
A few assumptions are baked in. They are worth naming.
You have an input pipeline that feeds the Friday-night queue. RSS feeds, newsletters (TLDR AI / The Batch / DIGEST / Alpha Signal), author whitelists, X / Bluesky lists. Without an input pipeline you have no queue to triage.
You have a reference manager you actually maintain. Zotero, Notion, Obsidian, plain markdown. The format matters less than your willingness to use it. Triage without a place to put the one-line notes is incomplete.
You are willing to deep-read 1-2 papers per week. If you are not, the workflow produces nothing useful — the triage exists to feed the deep-read habit.
If any of those three are missing, fix them first. The workflow does not work in isolation.
Where DIGEST fits in step 1
(Disclosure: we build DIGEST.) The Friday-night intake step is faster if the input pipeline does some of the triage signal for you. DIGEST's Quick Scan profile is shaped exactly for this — short summaries highlighting the contribution, organized per-category, delivered daily. By Friday you have a digest history that is already roughly priority-ordered.
This is not the only way to feed the queue. Author-whitelist RSS feeds work. Editorial newsletters work. DIGEST is one option if "auto-summarized arXiv per reader profile" matches your reading shape. The how-to-use page has a Quick Scan example.
The workflow is what matters, not the input source. If you are already maintaining a working pipeline, do not switch — the cost of switching pipelines usually exceeds the marginal gain.
Workflow audited against real-researcher feedback in 2026-Q2. Reviewed quarterly.