I Would Rather Quit NLP Than Read Another Paper Like This: How LLMs Are Mass-Producing the 'Rather Than' Pattern

An Increasingly Irritating Sentence Pattern

Have you noticed something when reading papers lately: pages full of 'X rather than Y' constructions?

• "We propose a new method, rather than simply scaling up the model"
• "Our method focuses on efficiency, rather than pursuing absolute accuracy"
• "We use end-to-end training, rather than relying on manual feature engineering"

One or two papers are fine, but when every paper is stuffed with this 'definitional negation'—first saying what something is not, then what it is—you start getting annoyed.

The title of this paper is itself a complaint: "I would rather quit NLP than read another paper like this".

The authors are not joking. They are serious. Using data, they demonstrate one thing: LLMs are turning the 'rather than' construction into a standard feature of NLP papers, and the source of this trend is post-training preference data.

The Data: 'Rather Than' Really Is Increasing

Q1: Has frequency actually increased?

The authors compared ACL papers from 2019 (pre-LLM ubiquity) and 2026 (post-LLM ubiquity):

• 2019: an average of 2.3 occurrences of "rather than" per paper
• 2026: an average of 6.8 per paper—a 3x increase

More importantly, LLM-generated papers (using 2019 and 2026 titles/abstracts to have LLMs write full papers) show even higher frequencies:

• LLM-written 2019-style papers: an average of 8.1 per paper
• LLM-written 2026-style papers: an average of 12.4 per paper

Human authors are converging toward LLM style, and LLMs push that style to even greater extremes. This is a positive feedback loop.

Q2: Do readers actually find it annoying?

Two senior NLP researchers (who routinely review for *ACL conferences) blindly annotated each "rather than" usage as "reasonable" or "annoying". Results:

• 2019 papers: almost no usages annotated as "annoying"
• 2026 papers: about 50% of papers contained at least one usage that annoyed the annotators
• LLM-generated papers: an even higher proportion of annoying usages

This shows: not all "rather than" usages are annoying, but 2026 papers do contain a large number of irritating ones.

Q3: What kinds of usage are annoying?

The annotators identified three characteristics of annoying usages:

1. Belittling the negated option: not just saying "we use A instead of B", but implying B is bad. "We focus on efficiency, rather than chasing unrealistic absolute accuracy"—the latter is framed as not worth pursuing. 2. Attacking strawmen: the negated option is often not something anyone actually does. "We use end-to-end training, rather than relying on outdated manual feature engineering"—but who still uses manual feature engineering in 2026? 3. Providing no informational value: negating an option doesn't increase the paper's precision. "Our method focuses on efficiency, rather than other aspects"—this says nothing.

Q4: Preference data is rewarding this pattern

Here comes the key finding. The authors examined open-source models' preference data (pairwise comparisons used for RLHF/DPO training) and found:

• Preference annotators tend to select responses containing "rather than"
• Open-source reward models score "rather than"-containing responses higher
• An instruction to "be honest" actually promotes this construction—because "rather than" looks like it's clarifying scope, appearing more precise

This explains why LLMs increasingly love this construction: post-training rewards it.

Why Does This Matter?

The grammatical version of "length bias"

Previous research found that LLM post-training has a "length bias"—longer responses score higher. This paper discovers a "contrastive-construction bias"—responses containing "rather than" score higher.

Both share the same essence: post-training optimizes certain surface features rather than genuine quality. Length bias makes responses longer; contrastive-construction bias makes responses more "contrastive"—whether or not the contrast is meaningful.

The "surface precision" trap

The "rather than" construction is likely rewarded by preference data because it looks more precise. "We use A rather than B" appears to delimit scope, suggesting clear thinking. But when the construction is overused—every sentence gets a "rather than"—it becomes an illusion of precision rather than actual precision.

This is highly related to the concept of the "rationalization shell": LLMs learn to wrap content in constructions that look precise, but the constructions themselves add no information. "Rather than" is the rationalization shell of NLP papers—it makes papers look more rigorous while actually being surface-level polish.

"Correcting surface forms is futile"

The paper ends with a warning: simply banning "rather than" won't solve the problem. Because post-training rewards the function of "contrastive framing", not the word "rather than". Ban "rather than" and models will switch to "instead of"; ban "instead of" and they'll switch to "not merely...but also".

This mirrors length-bias governance: punish long responses and models will produce "dense but uninformative" responses. The root of the problem lies not in surface form but in the post-training reward signal.

Deeper Implications

AI writing style "fingerprints" evolve

"Rather than" is just the currently identified fingerprint. Earlier research found ChatGPT favoring words like "delve", "tapestry", and "navigate". These words' frequencies spiked in 2023–2024, then began declining once identified.

This shows AI writing fingerprints are not static—it's a cat-and-mouse game with detectors. Once a detector identifies a fingerprint, post-training adjusts to the next unidentified one. "Rather than" is just the one identified in the current round.

Systematic bias in preference data

The more fundamental issue: preference data itself is biased. Human annotators doing pairwise comparisons tend to pick responses that "look more professional", and the "rather than" construction happens to trigger a "professionalism" heuristic.

This means: RLHF doesn't just optimize helpfulness—it also optimizes "surface features that look professional". Once these surface features are over-optimized, they become the "AI flavor" that irritates readers.

Implications for academic writing

If you're writing papers, this paper offers three practical suggestions:

1. Check your "rather than"s: for each one, ask "is anyone actually doing the negated option?" If not, delete it. 2. Don't define contributions through negation: saying "we did X" is enough; you don't need "rather than Y" to emphasize X's importance. 3. Beware the hidden influence of LLM-assisted writing: even if you don't have LLMs write full papers, LLM revision suggestions can push your phrasing toward "rather than".

Honest Assessment

1. Limited sample size: annotation was done by only two researchers, with considerable subjectivity; inter-annotator agreement is not reported in detail. 2. Only "rather than" studied: the frequency and annoyance level of other contrastive constructions ("instead of", "not merely", etc.) are not systematically compared. 3. Incomplete causal chain: the paper shows preference data rewards "rather than", but doesn't directly prove that preference data is the sole cause of the rise in NLP papers—human authors may also be imitating other humans. 4. "Annoying" is subjective: some readers may feel that well-used "rather than" constructions genuinely clarify scope. The paper acknowledges "legitimate" and "annoying" are not simple opposites.

But as a "problem-raising" paper, it is already powerful enough: it uses data to demonstrate a phenomenon many people vaguely sensed but nobody had systematically studied.

Conclusion

This paper itself demonstrates how to avoid the "rather than" trap—its title uses "rather than", but that's ironic self-reference. The body of the paper contains almost no gratuitous "rather than" usages.

The best scientific writing doesn't say "we did X rather than Y"; it says "we did X, because Z". Reasons are more persuasive than contrasts.

As LLMs increasingly participate in academic writing, we need not just AI-detection tools, but an understanding of what post-training is systematically changing. "Rather than" is just the tip of the iceberg—beneath the surface lies an entire reward bias in preference data.

---

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

Code: github.com/olzama/rather-than-public