NLP researcher probing how LLM persona representations are measured and manipulated in dialogue models.
2
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-16) and cover the author's entire public record, not just the papers listed here. Reproducibility signal comes from Evolve's own reviews.
Research profile
Output—
Not yet scored: publication-volume data is insufficient. This axis would measure the number, type, and venue spread of outputs over time; only two cataloged arXiv papers exist, so no output level is reported.
Impact—
Not yet scored: citation and influence data are unavailable. This axis would measure citations, uptake, and scholarly influence of the work; with no citation record provided, no impact estimate is offered here.
Focus0/100
Focus scores 0, the narrowest end of the scale, and the raw field distribution explains why: all cataloged work sits inside one subfield, with two nearly identical labels — 'natural language processing (LLM interpretability)' and 'natural language processing — LLM interpretability / representation analysis' — each weighted 0.5. Both 2026 papers target persona representations in dialogue LLMs, so the record is highly specialized rather than broad, though only two papers back this reading.
Collaboration41/100
Collaboration scores 41, a moderate level. The raw count shows 7 distinct coauthors across the cataloged papers, with the same names recurring — Huimin Han, Yu Sun, Cong Feng, and Guangming Lu appear repeatedly in the frequent-coauthor list, suggesting a small stable working group. Lu sits in middle author positions on both papers and is not corresponding author, consistent with contributing within a team rather than leading it.
Activity—
Not yet scored: longitudinal activity data is insufficient. This axis would measure publication cadence and sustained engagement over time; the visible record contains only two papers dated 2026, too thin to characterize an activity pattern.
Reproducibility55/100
Reproducibility scores 55, computed as the mean over 2 papers (reproMean 55, reproPapers 2), both rated medium readiness. This aligns with Evolve's verdicts: the cosine-similarity study is described as careful and well-controlled with PCA controls, while the K/V-cache paper candidly states limits but rests on one model/pair and a proxy behavioral measure. Both interventions are artifact-heavy but conceptually re-runnable.
Research summary
Mengyin Lu is an early-stage researcher in natural language processing, concentrated on interpretability and representation analysis of large language models, especially persona-related structure in dialogue-conditioned systems. The visible record consists of two arXiv papers, both dated 2026, in which Lu holds a middle author position and is not corresponding author. The first study argues that cosine similarity can underestimate linearly decodable persona structure in dialogue models, supporting the claim with supervised subspace projection and a PCA control; Evolve describes it as a careful, well-controlled empirical study with modest novelty. The second uses K/V-cache interventions — mid-layer replacement in Llama-3.1-8B — to dissociate representation alignment from persona expression in decoder-only models, candidly reporting limits, though Evolve notes the headline dissociation is partly by construction and the behavioral measure is a disfluency-marker proxy. Bibliometrically, the field distribution sits entirely within LLM interpretability (two near-identical labels weighted 0.5 each), yielding a focus score of 0. Collaboration scores 41, with 7 distinct coauthors and a recurring core of Huimin Han, Yu Sun, Cong Feng, and Guangming Lu. Reproducibility averages 55 across 2 papers, both rated medium readiness. Output, impact, and activity axes remain unscored for lack of sufficient underlying data.
Collaboration pattern
Mengyin Lu works within a small, recurring team: 7 distinct coauthors appear across the two cataloged papers, and Huimin Han, Yu Sun, Cong Feng, and Guangming Lu repeat in the frequent-coauthor list, suggesting a stable working group. Lu holds middle author positions on both papers without corresponding authorship, consistent with contributing to experiments and analysis rather than directing the agenda. The collaboration score of 41 reflects this moderate, stable pattern.
Research keywords
Highlights
Showing cosine similarity can miss linearly accessible persona structure
K/V-cache interventions that separate representation alignment from persona expression
Timeline
- 2026Published an arXiv study arguing that cosine similarity fails to reflect linearly accessible persona structure in dialogue models, with supervised subspace projection and PCA controls.
- 2026Published an arXiv study using mid-layer K/V-cache interventions in Llama-3.1-8B to dissociate representation alignment from persona expression.
- 2026Both cataloged outputs landed within LLM interpretability / representation analysis, establishing a focused early-stage profile whose two papers both carry medium reproducibility readiness.
Papers in the Evolve catalog
Frequent co-authors
About this profile
This profile is AI-structured by Evolve.Pub from public bibliographic data (OpenAlex) and Evolve's own paper reviews. It is not written or endorsed by the researcher unless the profile is claimed. Radar dimensions are computed from real bibliometrics with fixed formulas — raw figures are shown alongside; missing dimensions mean insufficient data, never zero. Errors? Claim the profile to correct or hide it.
- Bibliometric dimensions are computed from OpenAlex-derived data only; output, impact, and activity axes lack sufficient data and are left unscored.
- Profile is name-resolved from public bibliographic data; no ORCID, institution, or country is verified, so identity confidence is limited.
- Coverage is limited to the two arXiv papers reviewed by Evolve; other publications by this author may exist outside the catalog.
- The reproducibility score rests on a small base of 2 papers, both medium readiness, so it should be read as provisional.