What I learned watching people read papers

I sat in on 14 sessions of researchers reading papers in their own field. Not papers I picked — papers they picked, the night before. I took notes on where they paused, scrolled back, switched tabs, gave up.

The single biggest finding

Almost nobody reads a paper linearly. The dominant pattern was:

1. Abstract
2. Figures, in the order they appeared visually salient (not numerical order)
3. The first paragraph of the discussion section
4. Then back to methodology — but only the part that explained one specific figure

This is well-established in reading-comprehension literature; it was new to me, and surprisingly hard to design around.

What “reading order” implies for tools

If you’re building tooling for researchers reading papers — annotation tools, summarisers, comprehension aids — the implication is harsh:

• The figure is the primary unit of attention. Pretending the paragraph is is a polite fiction.
• A summariser that runs over the body text first is fighting the reader’s actual workflow. Run over captions and the first discussion paragraph first.
• A “smart highlight” feature must survive jumps — if I highlight in section 4 then jump to figure 2, the highlight needs to still be there when I come back.

What I’m doing differently now

When I prototype tools for this audience, I now start from the figures. Caption first. Body last. The reading-order data was an annoying surprise; now it’s a useful constraint.

The next study I want to run: what does this look like with eye-tracking, not screen recording? My hunch is that the figure attention is even more dominant than self-report admits.