How to Write a Results Section That Lets the Data Speak

1. What the Results Section Is — and Isn't

The results section has one job: report what you observed, accurately and without interpretation.

It states the facts your study produced. The discussion section is where those facts acquire

meaning. New authors blur the two, smuggling "this demonstrates that our method is superior" into

results — and reviewers notice immediately, because it reads as if you are afraid the numbers

cannot speak for themselves.

Keep the line sharp. "Accuracy rose from 81% to 89% (p < 0.01)" belongs in results. "This 8-point

gain shows our approach generalizes better" belongs in discussion. The discipline is not pedantry:

a reviewer who can separate your observations from your interpretation can evaluate each on its own,

and that separation is exactly what makes your claims credible rather than promotional.

2. Organize Around Claims, Not Chronology

Do not narrate your experiments in the order you ran them. Organize the section around the

questions your paper set out to answer, in the order a reader cares about them. Each subsection

should correspond to one claim or research question, open with the headline finding, and then

present the evidence that supports it. The reader should know the answer before they wade into the

numbers.

This mirrors the structure of your introduction: if you promised three contributions, the reader

expects three blocks of results, in the same order. A results section that wanders through

preliminary experiments, failed pilots, and tangents forces the reviewer to assemble the argument

themselves — and a reviewer doing your job for you is a reviewer in a bad mood. Put exploratory or

secondary analyses in an appendix and reference them.

3. Let Figures and Tables Carry the Evidence

The figure or table is the evidence; the text is the caption that tells the reader where to look.

Your worst move is to transcribe a table into prose — "Method A scored 0.84, Method B scored 0.79,

Method C scored 0.71..." — which wastes the reader's attention and duplicates what the table already

shows better. Instead, the text should state the pattern and direct the eye: "Our method outperforms

all baselines across every dataset (Table 2), with the largest margin on the low-resource split."

Every figure and table must be readable on its own, because reviewers often look at them before

reading a word. Label axes with units, define every symbol in the caption, and make the caption

state the takeaway, not just the contents. If a reader cannot understand Figure 3 without hunting

through the body text, the figure is not finished. (For the full treatment, see the guide on making

figures reviewers trust, linked below.)

4. Report Numbers Honestly — Effect Size, Not Just Significance

A p-value tells you whether an effect is likely real; it tells you nothing about whether the effect

is large enough to matter. Report both. "Significant" without an effect size is a common way to

oversell a trivial difference, and an alert reviewer will ask for the magnitude you omitted. Give

the actual numbers — means, differences, confidence intervals — and let the reader judge the size.

Always report the uncertainty around your central estimate: standard deviation, confidence interval,

or standard error, plus how many runs or samples the estimate rests on. A single number with no

spread is an anecdote. If you claim one method beats another, the comparison must account for the

noise; a 0.3-point gain with a 2-point standard deviation is not a result, and saying so yourself is

far stronger than letting a reviewer say it for you.

5. Handle Negative and Unexpected Results Head-On

Results that contradict your hypothesis are not failures to hide — they are findings to report.

Reviewers trust a paper more when it states clearly where the method did not help, where a baseline

won, or where an ablation removed something that turned out not to matter. Concealment is worse than

a weak result: if a reviewer suspects you buried an unfavorable comparison, they stop trusting every

favorable one.

Report the unexpected result plainly in the results section and save the explanation for the

discussion. "Performance dropped on the out-of-domain set (Table 4)" is a result; "we attribute this

to distribution shift" is discussion. Resist the urge to explain away a bad number on the spot —

doing so signals defensiveness, and it crowds the results section with the interpretation that

belongs one section later.

6. Common Mistakes and a Final Self-Check

The recurring failures are predictable: interpreting instead of reporting, restating tables in

sentences, leaning on p-values without effect sizes, hiding the runs that went badly, and presenting

experiments in lab-notebook order instead of argument order. Each one is cheap to fix before

submission and expensive to fix after a reviewer has flagged it.

Before you submit, read the results section as if you were a skeptic who wants to reject the paper.

For every claim, ask: is the evidence here, is the number honest, and is the interpretation kept out?

Then hand it to a labmate and ask them to state your main finding after reading only the figures and

their captions. If they get it right, the section works. If they hesitate, the evidence is not

speaking clearly yet.