Style over Substance? Linguistic Bias in LLM-based Peer Review Copy title

AI summary

Overview

Research area: Natural Language Processing — specifically LLM-as-a-judge evaluation, automated peer review, and bias/fairness analysis of language models.

Technical level: Intermediate. The paper is readable without deep technical background — the core idea is a controlled text-editing experiment — but it assumes familiarity with peer-review scoring conventions and basic statistical reporting (sign tests, preference ratios, standard deviations).

Scope: A controlled, content-preserving study of whether LLM reviewers change their scores and their written arguments when a paper's English proficiency style or its expressed confidence is altered, tested on 504 ICLR 2026 submissions with six LLM reviewers.

What This Paper Is About

Peer-review workloads at major AI conferences have grown sharply, and LLMs are increasingly used both by individual reviewers and by conference organizations to help produce reviews. The risk is that an LLM reviewer may be swayed by how a paper is written rather than what it says — for example, by fluent native-sounding English or by confident-sounding claims. The authors build a framework that rewrites real papers along two linguistic dimensions while keeping the science identical, then check whether LLM reviewers' scores and arguments shift.

Key Contributions

1. A controlled stress-testing framework for LLM reviewers. The authors create a pipeline that rewrites each paper along one linguistic dimension at a time in a content-preserving way, validates that the science is unchanged, and then compares reviews of the paired versions of the same paper.

2. Literature-grounded rewriting prompts for two dimensions of "epistemic authority." Rather than asking a model to "sound native" or "sound confident," the prompts are derived from specific linguistic findings — lexical diversity and sophistication, lexical bundles, syntactic variation, epistemic expressions, hedges and boosters — and the resulting rewrites are checked with established measures (MTLD, Advanced Guiraud, Gunning Fog, boosters and hedges per 1,000 words).

3. A large-scale empirical comparison across six LLM reviewers. Experiments cover 504 papers and reviewers from different families and capability levels, measuring score preferences, decision flips across the acceptance boundary, and the qualitative arguments raised in reviews.

4. Evidence that the framework can detect both the presence and the absence of a bias. When reviewers show no systematic preference along a dimension, the framework reports that; when the authors deliberately inject a style-weighting instruction into the reviewer prompt, the same framework exposes large, statistically significant preferences.

Main Findings

• Confidence dominates, proficiency barely registers. High-confidence rewrites are preferred over low-confidence rewrites by every reviewer tested, with preference ratios from 1.3× to 7.0×. The strongest are Claude Sonnet 4.6 (7.0×) and Gemma3 (6.0×). For native versus non-native rewrites, five of six reviewers favor the native version, but ratios stay between 1.0× and 1.3×.

• Average scores move in the same direction. Native versus non-native average ratings are nearly identical across models (for example 3.6/3.6 for Gpt-5.5, 4.7/4.7 for Gemini 3.5 Flash, 8.0/8.0 for Qwen3 VL), while high-confidence versus low-confidence pairs differ consistently (for example 4.6/4.2 for Claude Sonnet 4.6 and 7.2/6.7 for Gemma3).

• Confidence flips acceptance decisions. Using 4.5 as the acceptance threshold, Claude Sonnet 4.6 accepts only the high-confidence version in 21.0% of reviews and only the low-confidence version in 2.8%. Decision flips for the proficiency dimension are largely balanced, consistent with the weak score preference. Flips are rare for Gemma3 and Qwen3 VL because their average scores sit well above the threshold.

• The content really is preserved. Semantic similarity between rewrites and originals ranges from 0.997 to 0.999, and TF-IDF lexical similarity from 0.992 to 0.998 — close to the maximum of 1.

• The linguistic manipulations land as intended. MTLD rises from 60.8 (original) to 63.9 for native rewrites and falls to 58.0 for non-native; Advanced Guiraud goes from 7.7 to 8.4 (native) and 7.2 (non-native); Gunning Fog from 17.0 to 17.4 and 16.8. High-confidence rewrites contain 3.2 boosters and 0.8 hedges per 1,000 words, versus 0.7 boosters and 13.6 hedges for low-confidence rewrites.

• Confidence beats proficiency when both are combined. In a 126-paper combined experiment, native plus high-confidence language is preferred over non-native plus low-confidence language by up to 9.1× (Gemma3). Reversing the confidence dimension reverses the picture entirely: all reviewers then favor the non-native but high-confidence rewrite (ratios 0.3×, 0.2×, and 0.6× for Gemma3, Qwen3 VL, and Glm4.7).

• A deliberately style-biased reviewer is caught. Adding a general instruction to weight "linguistic style and overall presentation" produces native-versus-non-native preference ratios of 6.8 (Gpt-5.5), 4.3 (Claude Sonnet 4.6), and 2.2 (Gemini 3.5 Flash), all significant. This shows the framework can surface a bias the default reviewers did not display.

• The arguments change, not just the numbers. Using 16 coarse-grained aspects from an existing taxonomy, native rewrites lead to Presentation being criticized 2.4 percentage points less often and praised 3.0 points more often, a net polarity increase of 0.16; for high-confidence rewrites the Presentation gain is 0.29. Shifts in perceived Presentation do not correlate with score changes (r = 0.04), but the overall rating correlates strongly with the share of positive aspect mentions (r = 0.81). For high-confidence rewrites, Methodology, Theory, and Contribution are also perceived more positively.

• Reviewer variability differs sharply and must be accounted for. Pooled within-paper standard deviations are about 0.5 rating points for Gpt-5.5, Claude Sonnet 4.6, Gemma3, and Qwen3 VL, but 0.85 for Glm4.7 and 1.24 for Gemini 3.5 Flash.

Methodology in Plain English

The core problem is that you cannot simply compare real papers written in different styles, because they also differ in their science. So the authors hold the science fixed and change only the writing.

Step 1 — Rewrite. Starting from the original paper, they generate a paired version along one dimension: native-associated versus non-native-associated English, or high-confidence versus low-confidence phrasing. The prompts are not vague style requests; they encode specific linguistic properties documented in prior linguistics research, and they come with shared "content integrity rules" that require preserving length, structure, and technical content, and that forbid introducing grammatical errors or typos. Crucially, "low proficiency" does not mean broken English — both directions are plausible, publication-quality academic prose. A single model, Claude Opus 5, does all the rewriting so that reviewer differences are not confounded by the rewriter.

Step 2 — Check the rewrites. Two things must be verified: that the content survived, and that the intended style actually appeared. Content is checked with dense embeddings (BAAI/bge-large-en-v1.5) computed over passages rather than whole papers, plus TF-IDF cosine similarity on unigrams and bigrams to confirm technical terminology is intact. Style is checked with MTLD (lexical diversity), Advanced Guiraud (lexical sophistication), and Gunning Fog (textual complexity) for the proficiency dimension, and with booster and hedge frequencies per 1,000 words for the confidence dimension.

Step 3 — Review and compare. Each version goes to an LLM reviewer one at a time, using a prompt derived from standard conference reviewing instructions. The reviewer returns strengths, weaknesses, suggestions, sub-scores for soundness, significance, evidence, and clarity on a 1–4 scale, a confidence score on a 1–5 scale, and an overall rating on a 1–10 scale. The authors then compare the reviews of the paired versions of the same paper, using two-sided sign tests and preference ratios, and count how often the pair straddles the acceptance threshold.

Step 4 — Go beyond scores. Because review arguments matter as much as numbers, an LLM annotator (Gemma4, 26B) tags positive and negative mentions across 16 aspects, with each tag verified by substring matching against the review text. Net polarity per aspect is then compared within paper across rewrites.

Data and models. 504 papers are sampled from the ICLR 2026 reviewing cycle, uniformly across 21 primary areas, with 12 accepted and 12 rejected papers per area, all anonymized to simulate double-blind review. Reviewers are Gpt-5.5, Claude Sonnet 4.6, and Gemini 3.5 Flash (proprietary) plus Gemma3 (27B), Qwen3 VL (235B), and Glm4.7 (358B) (open-weight), all with a knowledge cutoff before 2026 so that knowledge of actual ICLR 2026 decisions cannot confound results. Reviewer stochasticity is measured separately: 10 papers, 10 repeated reviews each.

Why This Matters

Impact on research. If two versions of the same paper — identical in scientific content — receive different scores, different accept/reject outcomes, and different criticisms purely because one sounds more confident, then LLM review scores are not clean measures of scientific quality. This matters because LLM reviewing is being scaled up at exactly the moment submission volumes are exploding, and because linguistic style correlates with authors' language backgrounds, turning a stylistic artifact into a fairness problem. The paper also delivers a reusable evaluation instrument: a way to compare reviewer configurations by how stable they are under controlled linguistic perturbation, rather than by raw scores that the authors show are not directly comparable across models.

Real-world applications:

• Conference and journal review pipelines: organizations piloting LLM-assisted or LLM-generated reviews can run content-preserving rewrite tests before deployment to check whether confidence or fluency inflates scores.
• Model and vendor selection: the framework gives a concrete stability criterion for choosing among reviewer models, alongside the variance measurements the paper reports.
• LLM-as-a-judge evaluation more broadly: the same native/non-native and high/low-confidence perturbations apply to any judging task where style might leak into quality estimates, such as grading, ranking, or content moderation appeals.
• Author-side and editorial tooling: writing-assistance systems and editorial checks can flag when hedging or boosting levels are being read as evidence strength rather than as rhetorical framing.

Industry relevance. Any organization deploying an LLM as a judge — not just academic publishers, but grant agencies, benchmark maintainers, and platform trust-and-safety teams — faces the same question: is the judge rating substance or style? The paper's two headline practical lessons are that expressed confidence is treated as a proxy for quality by every model tested, and that raw scores from different LLM reviewers are not interchangeable without calibration, given average-score gaps and standard deviations ranging from about 0.5 to 1.24 rating points.

Future Directions

• Extend the stress test to other linguistic and rhetorical dimensions. The authors explicitly name this as future work, having covered only language proficiency and expressed confidence here.
• Calibrate reviewer scores. The paper notes substantial differences in average scores and within-model variance and cautions against treating raw scores from individual LLM reviewers as comparable without calibration — how to perform that calibration is left open.
• Determine whether the proficiency dimension is genuinely weak or merely under-triggered. The deliberate style-weighting experiment shows a large native preference can be elicited (up to 6.8×), raising the question of what real deployment conditions would produce such sensitivity.
• Test whether the confidence effect is fixable. Since reviewers appear to treat confident phrasing as evidence of stronger science, an open question is whether prompt design, rubric changes, or training can decouple expressed confidence from perceived scientific credibility — the paper motivates such checks but does not evaluate remedies.

Target Audience

Most valuable to researchers and practitioners working on LLM-as-a-judge systems, automated peer review, and evaluation methodology; to conference program chairs and journal editors considering LLM assistance in reviewing; to fairness and bias researchers interested in how linguistic style encodes author background; and to anyone building high-stakes LLM evaluation pipelines who needs a concrete protocol for testing whether a judge is sensitive to presentation rather than content. Readers with a linguistics background will also find the operationalization of hedges, boosters, lexical diversity, and lexical sophistication directly useful.

Authors’ abstract

The number of paper submissions to major AI conferences has increased dramatically in recent years, substantially increasing the peer-review workload of researchers. At the same time, large language models (LLMs) are increasingly used to support peer review, both by individual reviewers and by conference organizations. While LLM capabilities have improved considerably, their evaluations may be influenced by linguistic characteristics that are unrelated to scientific content. In this work, we introduce a controlled experimental framework for studying such effects by systematically perturbing the linguistic style of research papers. Specifically, we rewrite papers to exhibit (i) linguistic features associated with native or non-native English writing, and (ii) high- or low-confidence linguistic features. To construct these interventions, we derive fine-grained rewriting instructions grounded in the linguistic literature and apply them using LLMs. Crucially, our interventions are designed to preserve the underlying scientific content. We then evaluate the resulting papers using different LLM reviewers. Our experiments expose a critical vulnerability in LLM-based peer review: reviewers systematically reward confident rhetoric independent of scientific content, raising serious concerns about the robustness of their assessments.