Skip to main content

16. Results

A results section is not "here are our numbers"; it is "here is what our numbers mean."

Results, not analysis. That said, this section reports what happened — the scores, the comparison, what to see in each table — but it stops short of explaining why. The why is Analysis. The two blur, and the cost is a results section clogged with speculation and an analysis section with nothing left to do. Keep results to observation — "X beats Y by 3.2 points, and the gap widens on long inputs" — and save the mechanism — "because Y's attention degrades past 512 tokens" — for the next chapter. Reporting a number is results; arguing about its cause is analysis.

The shape of a good results section​

  1. Lead with the headline result. Table 1 or Figure 2 should answer the central question of the paper. The text should tell the reader what to see.
  2. One claim per paragraph. Each paragraph should make exactly one comparison or observation, with a pointer to the relevant table cell or figure region.
  3. Be specific. "Our method outperforms the baselines" is not a result. "Our method achieves 78.4 F1, a 3.2-point improvement over the strongest baseline (Smith et al., 2023, 75.2 F1)" is.
  4. Surface non-results too. If you ran an experiment and the difference was within noise, say so. "We find no significant difference between X and Y" is a perfectly publishable result and reviewers respect it.

Presenting the numbers​

Your results table and figures carry the headline; reviewers read them before the prose. The essentials specific to a results table:

  • Bold the best per column. Do not make reviewers squint to find the winning row. If the difference is not significant, bold both.
  • Show variance. "78.4 ± 0.6" beats "78.4" when you have multiple seeds.
  • Captions are paragraphs, not labels. A caption should let the reader understand the table or figure without the body text. Example: "Results on MMLU. Higher is better. Best per column in bold; differences within 0.5 are not significant. Our method is the last row."

The full craft of figures and tables — vector graphics, legible fonts, colorblind-safe and grayscale-readable design, booktabs, alt text, and the tools for each — lives in Chapter 7. Treat it as required reading before you build Table 1.

Phrasebook: the moves of a results section​

A results section is built from a few recurring moves — state the headline, point the reader at the table, report a comparison precisely, surface a non-result, scope the claim. As in the Methods and Experimental Setup phrasebooks, most entries below are adaptable templates with [bracketed] slots — fill every bracket with a real number, and never report a comparison you did not run. A few are real, verified sentences from well-known papers and carry a citation. Two entries sit in the open under each move; expand Show more for the rest. The cardinal rule survives all of them: a result is a number, a named baseline, and — where it matters — a test, never "outperforms the baselines."

Stating the headline result (from real papers and templates)​

The one sentence that answers the paper's central question with a number attached. Lead with it; everything else in the section supports it.

  • "It obtains new state-of-the-art results on eleven natural language processing tasks, including pushing the GLUE score to 80.5 (7.7 point absolute improvement), MultiNLI accuracy to 86.7% (4.6% absolute improvement)…" — Note what actually carries this sentence: the exact scores and the parenthetical deltas, not the phrase "state-of-the-art." BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding (Devlin et al., NAACL 2019)
  • "On [benchmark], [our method] reaches [score], a [delta]-point improvement over the strongest baseline ([cite], [their score])." — The headline template: metric, number, delta, named baseline.
Show 5 more
  • "This reduces the effort for finding the most similar pair from 65 hours with BERT / RoBERTa to about 5 seconds with SBERT, while maintaining the accuracy from BERT." — A headline stated as a concrete before/after on the axis that matters (time), with the accuracy caveat attached so it does not read as a free lunch. Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks (Reimers & Gurevych, EMNLP 2019)
  • "Table [N] is the main result: [one-sentence reading of what it shows]." — Tell the reader the table's headline before they parse the cells.
  • "The central finding is [X]; the rest of this section breaks it down by [dimension]." — States the conclusion first, then the supporting structure.
  • "[Our method] is best on [N] of [M] [tasks], and within noise of the best on the rest." — An honest aggregate headline that does not hide the losses.
  • "Gains are largest on [setting] ([delta]) and smallest on [setting] ([delta])." — A headline that already points at where the method helps and where it does not.

Pointing the reader at the table or figure (templates)​

A number the reader cannot locate is a number they do not believe. Name the table, the column, the row.

  • "Table [N] reports [what]; the comparison that matters is the [column] column, [model] versus [baseline]." — Directs the eye to the exact cell that carries the claim.
  • "Figure [N] plots [y] against [x]; the gap between [line A] and [line B] is the effect we claim." — Tells the reader what to see, not just that the figure exists.
Show 5 more
  • "Higher is better throughout; the best result per column is in bold." — States the reading convention once, so every later table is legible.
  • "We report [score] ± [std] over [N] seeds; bold marks results not significantly worse than the best." — Folds variance and significance into the table's reading rule.
  • "The trend, not the individual numbers, is the point of Figure [N]: [the trend]." — Tells the reader which features of a plot to read and which to ignore.
  • "Rows are grouped by [factor]; within each group, only [the variable] changes." — Makes a table's controlled comparison visible.
  • "Full per-[task] numbers are in Table [N] (appendix); the main text reports the [average / headline]." — Points to the exhaustive table while keeping the body focused.

Reporting a comparison precisely (templates)​

The difference between a result and a vibe is a number, a named baseline, and a test.

  • "[Our method] improves over [baseline] by [delta] points ([metric]); the difference is significant under [test] (p < [value])." — Delta, named baseline, and significance in one sentence.
  • "The gain over [baseline] is [delta], but falls within seed-to-seed variance ([std]); we therefore do not claim an improvement here." — The honest version when a difference does not survive noise.
Show 5 more
  • "Relative to [baseline], [our method] trades [delta down on metric A] for [delta up on metric B]." — Reports a trade-off honestly rather than cherry-picking the favourable metric.
  • "The improvement holds across [conditions] but disappears under [condition], where [observation]." — States exactly where a result generalises and where it stops.
  • "[Baseline B], not [Baseline A], is the one to beat; we exceed it by [delta]." — Names the genuinely strongest competitor instead of the convenient one.
  • "Both methods are within [delta] on [metric]; we report a tie and look to [secondary axis] to separate them." — Calls a tie a tie.
  • "The effect size is [number]; with our test set of [N], this is [adequately / not] powered." — Reads the result through power, per Card et al.

Reporting a non-result or a loss (templates)​

A clean negative is a finding. A buried loss is a liability a reviewer will dig up.

  • "We find no significant difference between [X] and [Y] ([numbers], p = [value]); this is itself informative, because [why]." — Reports a null result as a finding, with the reason it matters.
  • "[Our method] underperforms on [dataset] ([number] vs. [baseline]); we surface it here and examine the cause in Analysis." — Puts the loss in the main table and hands the why to analysis.
Show 4 more
  • "Contrary to our expectation that [hypothesis], [X] did not improve [metric]." — Frames a negative as a finding against a stated prior, not an apology.
  • "The gain we expected from [component] does not materialise; the ablation in Section [N] localises why." — Honest about an unmet expectation, and points to the evidence.
  • "On [hard subset], all methods including ours fail ([numbers]); we treat this as an open problem, not a footnote." — Surfaces a shared failure as a contribution.
  • "[Result] holds only under [condition]; we report the negative case rather than narrow the claim silently." — Discloses the boundary instead of quietly restricting scope.

Scoping the claim (templates)​

Every result has a boundary. Name it before a reviewer does, and never let a two-setting experiment become "robust across all settings."

  • "These results hold for [the settings tested: list]; we make no claim beyond them." — States the scope explicitly, the antidote to over-generalisation.
  • "We tested [N] [settings]; 'general' would require more, so we make the narrower claim that [precise claim]." — Refuses the overclaim and substitutes the precise one.
Show 4 more
  • "The improvement is consistent but modest ([delta]); we present it as incremental, not transformative." — Right-sizes the claim to the size of the effect.
  • "Our evidence is [in-domain / English / single-architecture]; whether it transfers to [other] is open." — Names the dimension along which the result is untested.
  • "We observe [correlation]; we do not claim [causation], which would need [the controlled experiment]." — Keeps a correlational result from being read as causal.
  • "Results are reported on [test set]; we did not evaluate on [other] and so do not speak to it." — Bounds the claim to what was actually measured.

Common mistakes​

  • No headline. The reader has to assemble the story from twelve tables. Tell them what to see.
  • Hiding losses. If your method underperforms on one dataset, do not bury it on page 9. Surface it in the main results table and address it. Reviewers find honesty more persuasive than spin.
  • Statistical sleight of hand. "Improves by 0.3 points" with no test of significance and a single seed is not a finding. It is noise.
  • Over-claiming generality. "Our method is robust across all settings" when you tested two settings. Be precise about scope.

Five examples worth studying​

  1. T5 (Raffel et al., 2020). The headline table compares many models on many tasks; the surrounding text directs the eye.
  2. Chain-of-Thought Prompting (Wei et al., 2022). The result figures do enormous work — pay attention to how each one isolates a single claim.
  3. Mission: Impossible Language Models (Kallini et al., ACL 2024 Best Paper). Results presented as a series of clean, controlled comparisons that build a cumulative argument.
  4. With Little Power Comes Great Responsibility (Card et al., EMNLP 2020). A results section that is mostly about what we cannot conclude — and a beautiful example of presenting non-results well.
  5. Beyond Accuracy: Behavioral Testing of NLP Models with CheckList (Ribeiro et al., ACL 2020 Best Paper). Tables that surface failure modes cleanly.

Further reading​