If you follow ML research, you know the feeling: ICLR, NeurIPS, ICML and CVPR each drop thousands of papers, and your "to-read" list quietly turns into a graveyard. Here's the workflow that finally worked for me. It's less about reading more and more about deciding faster.
1. Go topic-first, not conference-first
Browsing an entire conference proceedings front to back is how you burn out. Instead I start from a topic and pull the relevant papers across conferences and years.
Lately I've been using Paper List for this. You pick a topic (say, Retrieval-Augmented Generation) and it shows the scope up front, e.g. a few hundred papers across ~13 conferences and 5 years, before you dive in. Seeing the size of the space first genuinely changes how you budget attention.
2. Triage with a 3-pass skim
For each candidate paper:
- Pass 1 (~10s): title + abstract. Does it address my exact question? If not, drop it.
- Pass 2 (~1min): figures and the main results table. A paper usually reveals its real contribution in Figure 1 faster than in the prose.
- Pass 3 (deep read): only the handful that survived passes 1 and 2.
Most papers never make it past Pass 1, and that is the point.
3. Keep a living shortlist, not a backlog
A backlog is guilt. A shortlist is a tool. I cap "actively reading" at ~10 papers; anything older rotates out. If a paper keeps resurfacing in my own work, it earns its slot back.
4. Read for the delta, not for completeness
You rarely need the whole paper. Ask one question: what does this do that the previous state of the art didn't, and at what cost? If you can state that delta in a single sentence, you've extracted most of the value.
The tooling matters less than the mindset. Your bottleneck isn't reading speed, it's decision cost. Anything that helps you decide what NOT to read, whether that's a topic-first explorer, a hard shortlist cap, or a skim protocol, buys back more time than reading faster ever will.
How do you triage papers? Curious what workflows other people have landed on.
Top comments (0)