Source-linked AI summary
Writing Style Similarity Reflects Academic Genealogy
Cameron Manzo
TL;DR
Authorship attribution systems often assume writing style is individually owned, despite limited verifiable evidence about mentors’ influence. This paper links curated academic genealogy to attribution embeddings and finds mentorship-related writing similarity causes attribution errors 11.0× more often than chance.
Problem
Existing measurements lack verifiable evidence of how much researchers’ writing style comes from mentors, despite risks of false attribution during peer review.
Method
The study links curated Mathematics Genealogy Project records with embeddings of sole-authored arXiv abstracts while controlling shared fields and institutions.
Results
11.0× more attribution errors than chance involve the true author’s advisors or academic siblings, while shared institution and field alone show negligible similarity.
Takeaways & Limitations
Mentorship leaves a measurable trace in writing style that can confuse authorship attribution systems at rates far above chance.
Takeaways & Limitations
Linguistic background contributes 18.1% of similarity, which the study cannot separate from mentorship.
Abstract
from arXiv · showhide
As authorship attribution systems are increasingly deployed to detect ghostwritten and AI-generated papers, their errors can support accusations against legitimate authors. These systems assume each author's style is their own. Researchers, however, study under advisors, and inherit their stylistic quirks. We build a corpus of arXiv authors with $\geq 2$ solo papers from the Mathematics Genealogy Project graph, giving $5{,}803$ total authors and $2{,}501$ ground-truth advisor-student pairings. Using embeddings from a fine-tuned model, advisors sit $39.9\%$ closer in cosine distance to their students than a random same-field author does. Two open encoders reproduce the effect at $12.6\%$ and $14.5\%$. \emph{Academic siblings}, two students of one advisor who may never have met, sit $30.4\%$ closer across $8{,}360$ pairs, even when they studied at different institutions. Pairs who share only an institution and a field show negligible similarity. Given a closed-set attribution task over the same corpus, the system's errors occur on the true author's advisors and academic siblings $11$ times more often than chance.
1 Introduction
The section motivates testing whether mentorship contributes to writing-style similarity, because advisor-attribution false positives can be difficult to disprove. It introduces a ground-truth corpus linking curated Mathematics Genealogy Project records to attribution embeddings while controlling for shared fields and institutions.
- Motivation: Advisor-attribution false positives can be difficult to disprove during peer review, motivating study of whether mentors influence researchers’ writing styles.The section frames systematic false accusations as dependent on how much writing style comes from mentors.
- Method: The study controls for shared fields and institutions, which can independently inflate stylistic similarity, and uses curated Mathematics Genealogy Project records for verifiable relationships.Existing measurements lacked verifiable advisor influence; the corpus addresses that limitation through the MGP graph.
- Contribution: The contribution links MGP records to authorship-attribution embeddings, creating 2,501 advisor-student pairs and 14,398 total sibling pairs.The sibling total includes 8,360 pairs across different institutions, and the analysis examines failures in closed-set attribution tasks.
2 Related Work
Prior work links scholarly style evolution, social convergence, and academic genealogy to advisor influence and career-level transmission. This paper extends those lines by testing whether style travels through curated genealogy records to academic siblings, challenging authorship-attribution models’ independence assumption.
- Writing style dynamics: 13.7M publications show scholarly style trajectories stabilizing around publication 13, with measurable advisor influence in early-career work.Advisors are inferred as a scholar’s most frequent co-author during their first three publication years.
- Writing style dynamics: Style-convergence theory provides a mechanism for advisor influence by showing speakers converge toward high-status interlocutors’ styles in online communities.The present work extends this line beyond direct advisor links to academic siblings.
- Academic genealogy: Academic-genealogy studies treat advisor lineages as channels transmitting research topics, productivity trajectories, and career-level properties.The paper asks whether style is transmitted through the same genealogy channel.
- Authorship attribution: Authorship-attribution systems identify authors through nearest-neighbor searches over author embeddings, implicitly assuming each author’s style is independent.LUAR and StyleDistance represent purpose-built and content-independent state-of-the-art approaches, respectively.
3 Data and Methods
The study constructs a pre-2024 corpus of sole-authored arXiv papers linked to Mathematics Genealogy Project records, then represents authors with pooled text embeddings and evaluates genealogical and institutional relationships. It validates relationship effects through within-field label shuffles and replicates them with two open authorship encoders.
- Corpus: 5,803 people had text and embeddings from 6,133 eligible Mathematics Genealogy Project records, each requiring at least two single-authored arXiv papers.The corpus was restricted to pre-2024 papers and used sole-author papers to reflect individual style.
- Embeddings: 1,536-dimensional chunk embeddings were mean-pooled to paper level and then author level, with cosine distance measuring similarity.The main tuned encoder fine-tunes a Mistral-based authorship-attribution model; its weights and training data are unreleased.
- Institution assignment: 3,609 people (59%) were assigned across 459 unique institutions using their PhD-granting school recorded by the Mathematics Genealogy Project.Institution assignment records where an author received their PhD, not where they currently work.
- Relationship sampling: 2,501 advisor-student pairs and 14,398 sibling pairs with embeddings for both members formed the analysis samples after exact-name resolution and exclusions.Relationships came directly from the Mathematics Genealogy Project graph, while sibling pairs were split by whether members trained at the same institution.
- Statistical validation: 1,000 within-field relationship-label shuffles, or 10,000 for the lab-mate test, generated p-values from shuffled mean distances at or below observed values.The labels shuffled were advisor or institution.
- Open model replication: LUAR and StyleDistance provided open-model replication under authorship-representation and content-independent modeling objectives, respectively.The main analysis used the tuned encoder, while these two encoders tested replication with open models.
4 Results
Writing similarity tracks shared academic mentorship: advisors and academic siblings are substantially closer stylistically than random authors, while shared PhD institution alone shows negligible similarity. In closed-set attribution, systems disproportionately misidentify socially related authors, with such errors occurring 11.0× more often than chance.
- Institutional similarity: 3.2% below its permutation null for same-PhD-institution pairs under the tuned encoder, versus −1.0% for LUAR and −0.6% for StyleDistance.The tuned-encoder result uses 57,250 pairs among 3,083 people; on the normalized scale, it is 5.0% toward same-author similarity versus 61.7% for advisors.
- Advisor-student similarity: 39.9% closer than random under the tuned encoder, versus 12.6% for LUAR and 14.5% for StyleDistance; all three permutation tests give p < 0.001.On the normalized scale, tuned-encoder similarity reaches 61.7% toward same-person, versus 62.7% for StyleDistance.
- Academic siblings: 30.4% below null for siblings trained at different institutions, versus 34.2% below null for same-institution siblings, supporting a shared-mentor origin.Siblings often have no verified connection or temporal overlap, and both open models show the same trend.
- Attribution errors: 45.9% top-ranked true-author accuracy over 2,881 authors in the closed-set task produced 1,559 errors.Each author profile excluded one paper, candidates shared the author’s primary arXiv category, and ranking used cosine distance.
- Attribution errors: 8.02% of attribution errors selected an advisor, student, academic sibling, or same-PhD-institution author, versus 0.73% under uniform chance.This occurred 11.0× more often than chance, with p = 6.1 × 10^-85.
5 Discussion
The discussion interprets advisor and sibling style similarity as mentorship’s measurable imprint, while shared institution alone does not explain similarity. It also notes career-long stylistic change and limitations involving field diversity, topic, and linguistic background.
- Mentorship leaves a measurable trace in writing style, whereas sharing an institution alone does not.The resulting advisor and sibling similarity can confuse authorship attribution systems at rates greater than chance.
- Advisor similarity decays during an author’s career, consistent with stylistic imprinting and an independent voice emerging over time.The decline cannot be explained by students merely choosing initially similar-writing advisors.
- Academic siblings need not have met or published in the same field, making a common advisor the clearest explanation for their similar writing.These conditions rule out co-authorship or a shared lexicon as explanations.
- Limitations: The corpus is 71% math, 8% cs, 4% hep-th, and 2% stat, with the remainder spanning 40 other arXiv categories.This field distribution limits the study’s diversity because the Mathematics Genealogy Project originates from math.
- Limitations: 18.1% of variation is attributable to linguistic background, which the study cannot parse from mentorship.Topic confounds are bounded by StyleDistance’s content-independent training and a cross-field check.
A Normalized Effect-Size Scale
The paper uses a normalized effect-size scale anchored at 0 by an author’s own split-half distance and at 1 by a random same-field author. LUAR lacks a defined floor because it provides no per-document vector.
- 0 marks an author’s own split-half distance on the normalized effect-size scale.
- 1 marks the distance to a random same-field author.
- LUAR has no defined floor because it provides no per-document vector.
B Negative Control: Name-Collision False Matches
A surname-matched negative control shows modest stylistic similarity among unrelated same-field authors, indicating name collision as a limitation but not explaining the stronger advisor and sibling effects.
- Negative control: 18.1% closer: 2,712 unrelated same-field author pairs sharing a surname were stylistically closer than random pairs (p < 0.001).Surnames may proxy linguistic background, which shapes English writing style.
- Negative control: 43.2% for same-field advisors and 30.4% for different-institution siblings both exceeded the 18.1% surname effect.The larger effects indicate that surname overlap alone does not account for the advisor and sibling similarities.
- Negative control: The authors report surname-based similarity among unrelated pairs as a limitation of the study.The negative control demonstrates a residual association that can produce false matches.
C Cross-Field Robustness Check
Cross-field advisor pairs, who share no subfield, remain substantially closer than the null baseline, though less so than same-field pairs. This significant result indicates the similarity effect is not purely topic-driven.
- C Cross-Field Robustness Check: 26.7% below their null (p < 0.001), cross-field advisor pairs show a weaker but still sizable and significant effect than the 43.2% for same-field pairs.The 473 cross-field pairs share no subfield by definition, so the effect cannot be explained solely by topic overlap.
D Career-Phase Analysis
Students’ writing becomes less similar to their advisors over their careers: mean cosine distance rises from 0.0526 early to 0.0563 late, a +7.1% change. This pattern supports advisors imposing style on mentees rather than students initially selecting stylistically similar advisors.
- D Career-Phase Analysis: The analysis uses 1,286 advisor-student pairs whose students have at least four dated papers.Each student’s papers are sorted by arXiv date, with the earliest and latest halves pooled into early- and late-career profiles.
- D Career-Phase Analysis: 0.0526 early-career mean cosine distance rises to 0.0563 late-career, a +7.1% change.Distances are measured against the advisor’s profile.
- D Career-Phase Analysis: The increase refutes students choosing advisors who already write similarly and supports advisors implicitly imposing their style on mentees.