I would rather quit NLP than read another paper like this: The rise of antithesis in NLP papers

1 Introduction

• The paper situates its concern in reports that corrective framing is increasingly common and can irritate reviewers and other readers.It distinguishes this concern from the ordinary use of “rather than” to express alternatives and preferences.Source: §1 · p. 1
• The study asks whether “rather than” has grown more frequent, which uses annoy readers, what distinguishes those uses, and whether preference data reward the construction.The authors frame these as preliminary questions for a larger study involving psychologists and specialists in commercial-model post-training.Source: §1 · pp. 1–2
• The authors compare ACL 2019 papers, ACL-style arXiv 2026 papers, and GPT-generated papers, and blindly annotate sampled occurrences as legitimate or annoying.They also examine evaluative stance and straw-man alternatives, then test whether preference data and reward models favor the construction.Source: §1 · p. 2
• The construction is far more frequent in recent NLP papers and more frequent still in LLM-written papers.The authors report that almost only 2026 uses were found annoying, and about half of 2026 papers contained a use that annoyed a given annotator.Source: §1 · p. 2
• Annoying uses disparage rejected alternatives more, while preference-data raters and open reward models favor the construction.An instruction to be honest also promotes it; the authors conjecture that pairwise-preference post-training may produce this pattern.Source: §1 · p. 2

2.1 LLM-associated change in academic writing

• Corpus-scale studies identify LLM-associated lexical change, while direct comparisons report formulaicity, reduced variation, and shifts in vocabulary, formality, and sentiment.Some lexical markers later decline while others continue rising, making simple lexical detection unstable; these studies do not assign authorship to individual papers.Source: §2.1 · p. 2

2.2 Contrastive constructions

• Linguistic work shows that antithesis markers can serve distinct contrastive functions, motivating analysis of construction function rather than treating every occurrence as suspect.The closest cited LLM study reports broader corrective framing, including “rather than,” at roughly 12 times the human rate for GPT-6 Astra.Source: §2.2 · p. 2

2.3 Preference optimization and style

• Preference optimization can amplify features raters favor, including length, lists, bold text, links, emojis, and agreement with users’ views.Prior work describes reward models as imperfect proxies and discusses correcting exploited features one at a time; this paper asks whether a discourse construction is another such feature.Source: §2.3 · p. 3

3 Data

• The dataset combines ACL 2019 papers, ACL-style arXiv 2026 papers, and GPT-generated papers prompted with titles and abstracts from both years.The authors expect substantial LLM assistance in many 2026 papers but cannot determine how much an LLM wrote any particular paper.Source: §3 · p. 3
• ACL 2019 covers all ACL Anthology venues; the 2026 corpus keeps cs.CL submissions naming an ACL venue and confirms ACL style in source files.Language and text-quality filters exclude unusable English text, and rates use body text with references and appendices removed.Source: §3 · p. 3
• Body length rises by 67% from 2019 to 2026, but format differences explain much of the gap and excluding extra sections narrows the long-paper gap to 14%.Rates per 1,000 words are unaffected by length, while counts per paper used in the encounter model are affected.Source: §3 · p. 3
• For 807 arXiv 2026 papers with multiple versions, the authors collect first and latest versions; 793 pairs have usable body text in both.These pairs support analysis of how authors alter the construction during revision.Source: §3 · p. 3

4.1 Antithesis patterns

• The authors count “rather than” and related patterns per 1,000 body-text words, using regex heuristics and precision checks while leaving recall unestimated.They examine seven related constructions, but omit two sparse emphatic patterns from frequency results.Source: §4 · p. 3, §4.1 · pp. 4–5

4.2 Generating papers with LLMs

• Three OpenAI models generate ACL-style papers from titles and abstracts, then review and revise them; Tülu checkpoints and Claude models extend the pipeline.The extended runs cover the first 100 source papers and compare Tülu checkpoints at different post-training stages.Source: §4.2 · p. 5

4.3 Human annotation

• Two experienced NLP researchers blindly label a 1,000-item pool as annoying, legitimate, unsure, or garbled, with source and document hidden.The pool samples multiple corpora and batches; a separate GPT comparison set is analyzed separately because annotation sessions were spaced apart.Source: §4.3 · pp. 5–6

4.4 Characterizing uses

• A case-control sample of 121 arXiv 2026 uses tests whether annoying items more often reject straw men; separate stance ratings assess how authors present each alternative.Two LLM coders classify disavowals, and straw-man coding is conducted without annoyance labels, so the comparison tests association rather than overall prevalence.Source: §4.4 · pp. 6–7
• The authors test syntactic patterns and semantic similarity across annoying and legitimate uses, and compare perceived LLM authorship with annoyance judgments.The syntax scan uses parsed sentence features; authorship judgments use short passages, most outside the judged occurrence’s immediate context.Source: §4.4 · pp. 7–8

4.5 Preference data and honesty instructions

• Preference datasets are analyzed with conditional logistic models of chosen responses, controlling for pattern presence differences and response length; randomized honesty prompts test instruction effects.The honesty analysis compares each principle with helpfulness while controlling for source dataset, generating model, and log length.Source: §4.5 · p. 8
• Open reward models score original sentences against versions with the contrast clause deleted or replaced by a length-matched non-contrastive clause.This minimal-pair design isolates clause-level preferences more directly than comparisons between different responses, though deletion also removes content.Source: §4.5 · pp. 8–9

4.6 Analysis

• The analyses use tests suited to proportions and scores, adjust multiple-comparison tests, and distinguish preregistered predictions from exploratory analyses.Safeguards include corpus-blind annotation, checks of LLM judgments against human judgments, and random assignment of UltraFeedback system prompts for the honesty test.Source: §4.6 · p. 9

5.1 Frequency

• About seven times: “rather than” occurred at about seven times the 2019 rate in arXiv 2026 papers, and was even more frequent in GPT-5.6-sol and GPT-6-sol papers.The increase was concentrated in “rather than” and postposed “X, not Y”; “instead of” and “as opposed to” declined, while emphatic “not only/just/merely X but Y” rose modestly.Source: §5.1 · p. 9, Figure 1 · p. 9
• GPT-5.6-sol and GPT-6-sol used “rather than” in every paper, while Claude models did so in nearly every paper and Tülu 3 8B models in a minority.Revision did not reliably reduce use: GPT revisions changed rates in opposite directions, while Claude revisions varied by model; preference optimization or reinforcement learning with verifiable rewards did not raise rates beyond supervised fine-tuning.Source: §5.1 · p. 10

5.2 Reader judgments

• Annotators found almost no 2019 uses annoying, while about one in ten 2026 uses were annoying; about half of 2026 papers had a use annoying each annotator, though annotators seldom agreed which uses.Source: §5.2 · p. 10

5.3 What makes a use annoying

• Annoying uses presented the affirmed alternative above the rejected one more strongly than legitimate uses, and rejected alternatives were more often judged straw men, though coders disagreed widely.The stance contrast persisted after excluding inverted uses, although it was smaller; some syntactic patterns also associated with annoyance, but differed by annotator.Source: §5.3 · pp. 12, 13, 15
• Negative treatment of the rejected alternative was more common in arXiv 2026 than ACL 2019 uses and more common still in GPT papers, especially GPT-6-sol’s.Claude papers did not disparage rejected alternatives more than their source papers, showing that high construction frequency did not necessarily coincide with harsh stance.Source: §5.3 · pp. 13–14
• Disavowals were far more often annoying and nearly unrelated to stance; they were absent from ACL 2019 and GPT-4o uses, and more common in LLM-written papers and revisions.The association between disavowal and annoyance held for A1 in every set and for A2 in the GPT comparison set; the evidence does not establish why revisions contain more disavowals.Source: §5.3 · p. 14
• Annoying uses resembled one another in their rejected alternatives and sentences, while perceived LLM authorship did not consistently correspond to annoyance.Many annoying rejected alternatives described implausible positions or methodological flaws; annotators agreed more on passages seeming LLM-written than on which uses annoyed them.Source: §5.3 · p. 15
• Other tested factors—including LLM-style markers, nearby antithesis, and blatant straw men—showed no consistent association with annoyance; domain and paper-level rather than rate were associated for A2 only.A2’s annoying labels clustered by paper, whereas A1’s did not.Source: §5.3 · p. 16

5.4 Where the construction comes from

• Rather than raised response-choice odds in three of four preference datasets, most strongly for human-rated data; human raters rewarded it at least as much as LLM judges did.The effects were modest and controlled only for length; other response differences could account for some of the preference. Emphatic not only/just/merely X but Y raised odds in all four datasets.Source: §5.4 · p. 16, Table 23 · p. 16
• All four reward models preferred sentences with rather than clauses to versions with the clause deleted, and the original beat length-matched controls for every model.Much of the gain came from adding text; the original-versus-control difference persisted after accounting for sentiment. The comparison cannot isolate antithesis from the content of Y, and all four models are from one family.Source: §5.4 · p. 16, Table 24 · p. 17
• For three of four reward models, clauses in uses A1 labeled annoying earned more length-adjusted reward than clauses in legitimate uses; A2 showed no such difference.The models gave the most credit to clauses that annoyed at least one reader.Source: §5.4 · pp. 16–17, Table 25 · p. 17
• An honesty instruction modestly increased rather than use and shifted models toward correction or disavowal, while explicit confidence-level instructions increased none of the tested forms.The experiments used 2023 models on general assistant tasks; the effect on rather than was small, and whether an honesty training objective has the same effect remains untested.Source: §5.4 · p. 17

• Annoying uses favor the affirmed option and disparage the rejected one more than legitimate uses, a pattern robust across annotators and labeling batches.Source: §6 · p. 17
• Straw-man codes track annoyance closely for A1 but not A2; perceived LLM authorship tracks one annotator, while blind readers treat the construction itself as an LLM cue.Source: §6 · p. 17
• Post-training may reward disavowals as displays of calibration, though the evidence is indirect and suppressing rather than could shift the same function to another form.Source: §6 · pp. 17–18
• Open reward models score rather-than clauses above length-matched noncontrastive alternatives, but Tülu 3 8B checkpoints show no increase from supervised fine-tuning to later stages.Source: §6 · p. 18
• That checkpoint comparison is limited by one sample per stage and few instances; testing a model with a known post-training history that includes the habit is still needed.Source: §6 · p. 18
• Annotators seldom agree on individual annoying uses, yet aggregated judgments may converge as many uses encounter readers' distinct triggers.Source: §6 · p. 18
• Whether annoyance grows with exposure remains open; A1's impression of sensitization contrasts with flat rates over labeling order.Source: §6 · p. 18

• The rather than construction has become a marker of LLM-era NLP writing, though the authors expect this particular marker to be short-lived.Source: §7 · p. 18
• Annoyance correlates most robustly with evaluative stance and, for one annotator, with perceived likely LLM authorship, while stance and rate do not explain differences across model families.Source: §7 · pp. 18–19
• Human-signed papers have adopted the construction, moving literature style toward that of models, which will in turn learn from it; judgments should span many readers and whole texts.Source: §7 · p. 19

Limitations

• The study covers English scientific NLP papers; its arXiv comment-field heuristic may miss eligible papers, and regex recall is unmeasured.Some patterns overgenerate, especially in ACL 2019 PDF text.Source: Limitations · p. 19
• Annoyance judgments came from two author-annotators who knew the hypotheses and had previously labeled different items from the same papers.Both were experienced reviewers, but differences in their judgments cannot be separated from differences in experience.Source: Limitations · p. 19
• The disavowal measure relies on two LLM coders, has moderate agreement with A1, and its association with annoyance is weaker for A2.The straw-man test set came from one annotator’s labels, and two coders had already judged annoyance.Source: Limitations · p. 19
• Preference-data results are correlational, control only for length, and come from four general-assistant datasets unrelated to scientific writing.Reward-model tests used minimal pairs from one family, while post-training stages were tested only preliminarily on one open model.Source: Limitations · p. 19

A Corpus construction

• 4,863 of 4,943 attempted ACL Anthology papers yielded usable text for the 2019 corpus.The authors extracted plain text from PDFs with pdftotext.Source: A Corpus construction · pp. 21–22
• 1,841 confirmed ACL-style papers remained from 1,985 arXiv 2026 comment-field candidates after source inspection and exclusions.Twelve candidates failed to download, and 132 downloaded candidates did not use the ACL style.Source: A Corpus construction · p. 22
• After excluding documents with insufficient usable English text, the analysis uses 4,744 ACL 2019 papers, 1,821 arXiv 2026 papers, and 793 version-diff pairs.The filter excluded 119 ACL 2019 documents and 20 arXiv 2026 documents; the version-diff corpus also excludes 20 near-empty documents.Source: A Corpus construction · p. 22

B The residual length gap between ACL 2019 and arXiv 2026

• 67% body-length growth from 2019 to 2026 narrows to 36% among long papers and 14% after excluding selected sections.Citation and cross-reference density may explain part of the residual, but arXiv placeholder counts are not directly comparable with ACL 2019 counts.Source: B The residual length gap between ACL… · p. 22, Table 29 · p. 22

C Annotation instructions

• Annotators judged whether each highlighted use annoyed them and whether its rejected alternative was a straw man, using separate instructions for each judgment.The annoyance label depended on the annotator’s reaction; a straw man was an alternative no competent author would seriously adopt or defend.Source: C Annotation instructions · p. 23

D Prompts for the LLM-written papers

• The LLM-paper prompts supplied source titles and abstracts, prescribed ACL-style paper sections, and included review and revision stages.Revision instructions asked models to act on methodology and development suggestions without fabricating results.Source: D Prompts for the LLM-written papers · pp. 23–25

E Generation settings

• All models received the same prompts, while generation settings and access routes differed between OpenAI and Claude models.Claude Code did not allow temperature or top-p settings and added contextual lines to each request.Source: E Generation settings · p. 25

F Verbatim overlap with the source papers

• No rather-than sentence in the generated GPT, Tülu 3 8B, or Claude papers shared a 13-gram with its source paper.The authors report that the generated papers did not reproduce their source papers, despite possible exposure to ACL 2019 papers during model training.Source: F Verbatim overlap with the source… · p. 25

G Coding disavowal

• Disavowals decline a stronger or broader claim about the authors’ own work, whereas approach contrasts describe choices and finding contrasts assert what a finding is.The coding instructions classify uses into disavowal, approach contrast, finding contrast, or other.Source: G Coding disavowal · p. 25
• The coders agreed closely on disavowals, but their agreement with A1’s broader coding was moderate and did not improve with revised instructions or examples.A1’s coding included rejections of limiting characterizations of the authors’ work.Source: G Coding disavowal · p. 25
• Disavowals and evaluative stance were nearly unrelated; disavowals had a slightly smaller stance gap because the declined claim was not scored as disparaged.The authors report that neither property strongly tracked the other.Source: G Coding disavowal · pp. 25–26

H Revision of the Claude and Tülu papers

• Haiku 4.5 and Opus 5.5 add rather than in revision, while Sonnet 5.5’s rate does not change.Sonnet 5.5 and Opus 5.5 also add postposed “X, not Y”; Haiku 4.5 does not.Source: H Revision of the Claude and Tülu papers · p. 26
• Most sentences containing rather than retain it in revision, and additions usually go into sentences that had no antithesis form.When rather than is removed, it is usually not replaced by another antithesis form.Source: H Revision of the Claude and Tülu papers · p. 26
• Tülu 3 8B uses rather than too rarely for a sentence-level comparison, and its revisions raise neither construction.Source: H Revision of the Claude and Tülu papers · p. 26

I Revision histories

• 0.988 vs. 0.921 occurrences per 1,000 words: the latest versions exceed v1, with added rather than hunks outnumbering removals 1,156 to 576.The mean rate rises significantly, although about as many papers go up as down; the increase comes from a subset adding many instances. Other patterns are mostly flat or slightly down.Source: I Revision histories · p. 27, Table 38 · p. 27, Figure 2 · p. 28
• Surgical removals are a minority of phrase removals, so most edits do not clearly isolate a judgment that rather than itself was unnecessary.The GitHub history source is thin: most repositories are single uploads with no edit history.Source: I Revision histories · p. 28

J Annotation pool

• The ACL 2019–arXiv 2026 difference in annoying labels remains significant when reference and appendix items are excluded; version strata have similar annoying rates.The version-strata rates are lower than the main arXiv 2026 sample, but not significantly so.Source: J Annotation pool · p. 28

K Labels across batches

• Across batches, A1’s corpus difference and stance gap appear in the first batch alone, while the we-subject association appears only in the second.The stance gap does not differ between batches (interaction p = .82 for A1; p = .54 for A2).Source: K Labels across batches · p. 29

L Annotator drift

• Most repeated labels stay the same, but A2’s annoying rate rises between batches labeled on different days; random ordering makes stratum comparisons less exposed to this drift than absolute rates.A1’s rate is flat after the first fifth; A2’s rate is flat within each batch.Source: L Annotator drift · p. 29

M A5’s labels

• A5 labels 12 of 16 ACL 2019 items and 17 of 26 random arXiv 2026 items as annoying, similar shares across the two corpora.The comparison is not significant (Fisher’s exact test, p = .73).Source: M A5’s labels · p. 29

N Further correlates of annoyance

• Uses from papers outside core NLP are labeled annoying more often, significantly so only for A2; the predicted core-versus-other difference does not appear.The direction was not predicted, so the authors treat it as a hypothesis for new labels.Source: N Further correlates of annoyance · p. 29, Table 47 · p. 30
• Neither nearby antithesis nor the tested AI-writing tells is associated with item-level annoyance for either annotator; whole-paper rather than rate is associated only for A2 in an unplanned comparison.Sentence-initial emphatic adverbs and em dashes were too rare to test.Source: N Further correlates of annoyance · p. 29, Table 48 · p. 30
• A2’s annoying labels cluster within papers, but the apparent link between a straw man elsewhere and annoyance does not remain after adjustment for the item label and paper’s rather-than rate.Source: N Further correlates of annoyance · p. 30, Table 49 · p. 31
• No lexical term survives correction; the clearest shared straw-man examples are annoying to both annotators, while “X-ed rather than Y-ed” cases are mostly ordinary contrasts despite rising in share across corpora.The past-participle pattern rises further in GPT papers, especially after revision.Source: N Further correlates of annoyance · pp. 30–31, Table 52 · p. 31

O Correlates of annoyance in the GPT comparison set

• In the GPT comparison set, annoyance correlates with a we-subject pattern for A1 and passive clauses and past participles for A2; sentence-initial Rather than is too rare to test.Source: O Correlates of annoyance in the GPT… · pp. 31–32, Table 53 · p. 32

P Perceived LLM authorship

• Passages containing rather than are judged LLM-written about twice as often as passages without it, including by annotators who did not know the study’s hypotheses.The same effect appears in judgments by A3 and A5.Source: P Perceived LLM authorship · p. 32, Table 55 · p. 32

Q Checks of the stance ratings

• The stance-rating association with annoyance replicates in the pool sample using Claude ratings, but is weak in the Claude comparison set.A human judgment of favoring Y agrees with GPT-6-sol’s inverted rating 78% of the time (κ = 0.55); the Claude agents also order uses similarly to GPT-6-sol and replicate the pool association.Source: Q Checks of the stance ratings · pp. 32–33, Table 56 · p. 32, Table 59 · p. 33

R Straw-man coding: agreement and LLM coders

• Human coders vary widely in how many items they call straw men and agree little on which items; GPT-6 coders perform about as well as human coders.Human counts range from 7 to 42 among the same 151 items; LLM agreement with humans is κ = 0.14–0.17 with the definition alone and 0.27 or 0.22 with examples.Source: R Straw-man coding: agreement and LLM… · p. 33, Table 60 · p. 33

S Exposure effects in annotation

• A1 showed no net labeling-order shift, while A2 labeled more items annoying in batch two despite its larger share of arXiv 2026 items.A1 nevertheless perceived growing annoyance, and all of A1’s ACL 2019 annoying labels came from batch two; annotators may differ in susceptibility to exposure effects.Source: S Exposure effects in annotation · pp. 33–34
• Planned short annotation sessions sample judgments before heavy exposure, while varying prior item exposure would directly test exposure effects.Source: S Exposure effects in annotation · p. 34
• Claude’s higher revised-use annoyance rate persisted after controlling exposure measures; among 97 draft/revision pairs, only five were near-identical, and none was annoying.Neither session position nor the number of earlier uses from the same paper predicted annoyance.Source: S Exposure effects in annotation · p. 34, Table 61 · p. 34