Do Implicit Personalization and Explicit Styles Conflict? PsPLUG: A Lightweight Plug-in for Balancing Personalization and Style in Customized LLMs

Trait induction and control2026arXivApproved editorial review

Authors: Yutong Song, Jiang Wu, Shaofan Yuan, Chengze Shen, Jian Wang, Yu Wang, Nikil Dutt, Amir M. Rahmani

Keywords: Personalized generation, Style-conditioned generation, Soft prompts, Preference optimization, LaMP benchmark

Source: Open primary source (opens in a new tab)

8
Authors
8
Findings
10
Limitations
4
Evidence

Editorial summary

English

ArXiv v2 studies whether an explicit style instruction can erase user patterns that a personalized system is intended to preserve, a failure mode it calls personalization collapse. PsPLUG freezes a Qwen3 backbone and prepends three continuous embeddings: a shared system vector, a user vector, and a query-dependent vector. To build the user vector, Qwen3-8B summarizes ten history items with a task-specific prompt; BGE-base-en-v1.5 encodes that profile and a trainable MLP projects it into the LLM space. Training treats the real user response as preferred and an impersonal style-conditioned generation as rejected in a Bradley-Terry/DPO loss. At inference, alpha scales only the user vector.

Without style instructions, the paper evaluates six LaMP tasks. PsPLUG is best on 12 of 15 metrics, ties RAG on LaMP-7 METEOR, and trails OPPU on LaMP-2 accuracy and RAG on LaMP-5 METEOR. With four styles, warm, critical, concise, and elaborative, on LaMP-4, LaMP-5, and LaMP-7, the numerical values place it first or second in 22 of 24 ROUGE cells. It falls outside the top two on LaMP-5 concise ROUGE-L and LaMP-7 critical ROUGE-L. The table contains at least six incorrect best/second markings, although the over-80-percent claim still holds when ranks are recomputed. No repeated runs, intervals, or significance tests are reported; several differences are only 0.001 to 0.005.

The evidence does not cleanly separate persona from style. The positive response was written by the user without the newly imposed style instruction, while the negative follows that style; the contrast changes personalization and style compliance at the same time and cannot identify a pure residual. The style tables also call ROUGE against the original unstyled reference a personalization measure: obeying the requested style can reduce overlap, while ignoring it can increase overlap. Human and GPT-5.2 Pro evaluation is shown only for LaMP-7. The paper does not report how many items were annotated, who annotated them, inter-rater agreement, or raw ratings. Correlations use means aggregated over five systems and four styles, at most 20 points, rather than item-level judgments and have no uncertainty. They therefore support directional rank association, not claims that the judge is nearly indistinguishable from humans or measures true persona adherence. The 8B values in Table 4 also conflict with Table 2; for example, LaMP-5 METEOR is 0.321 versus 0.398.

Reproducibility is blocked. The anonymous code link in v2 has expired. A public homonymous repository contains five partial LaMP-7 scripts, but it cannot be unambiguously tied to the expired archive and includes no data, results, checkpoints, dependencies, judges, human evaluation, tests, or CI. Training saves nested checkpoint keys incompatible with the flat keys required by inference; alpha is not exposed; the DPO reference is another frozen PsPLUG policy copied at initialization rather than the equation's base model; and the other tasks and style sweeps are absent. The contribution supports compact-profile prompting as a plausible personalization direction and shows that explicit styles alter reported LaMP behavior. It does not establish mathematical disentanglement, a calibrated trade-off, measured efficiency, privacy, psychological personality, or end-to-end reproduction. Bibliographically, v2 remains a preprint and replaces v1's title, abstract, and author order.

Español

La versión arXiv v2 estudia si una instrucción explícita de estilo puede borrar los patrones de un usuario que un sistema personalizado intenta conservar, fenómeno que denomina personalization collapse. PsPLUG mantiene congelado un backbone Qwen3 y antepone tres embeddings continuos: un vector de sistema compartido, un vector de usuario y un vector dependiente de la consulta. Para construir el vector de usuario, Qwen3-8B resume diez elementos del historial con un prompt específico de la tarea; BGE-base-en-v1.5 codifica ese perfil y un MLP entrenable lo proyecta al espacio del LLM. El entrenamiento trata la respuesta real del usuario como preferida y una generación impersonal condicionada por el estilo como rechazada dentro de una pérdida Bradley-Terry/DPO. En inferencia, alpha escala solo el vector de usuario.

Sin instrucciones de estilo, el paper evalúa seis tareas de LaMP. PsPLUG obtiene el mejor valor en 12 de 15 métricas, empata con RAG en METEOR de LaMP-7 y queda por detrás de OPPU en accuracy de LaMP-2 y de RAG en METEOR de LaMP-5. Con cuatro estilos, warm, critical, concise y elaborative, sobre LaMP-4, LaMP-5 y LaMP-7, los números lo sitúan primero o segundo en 22 de 24 celdas ROUGE. Queda fuera de los dos primeros en LaMP-5 concise ROUGE-L y LaMP-7 critical ROUGE-L. La tabla contiene al menos seis marcas de mejor/segundo incorrectas, aunque la afirmación de superar el 80% sigue cumpliéndose al recalcular los rangos. No se publican repeticiones, intervalos ni pruebas de significación; varias diferencias son de 0,001 a 0,005.

La evidencia no separa limpiamente persona y estilo. La respuesta positiva fue escrita por el usuario sin la instrucción de estilo nueva, mientras la negativa sí sigue ese estilo; el contraste cambia a la vez personalización y cumplimiento estilístico y no identifica un residual puro. Además, las tablas de estilo llaman personalización a ROUGE contra la referencia original no estilizada: obedecer el estilo puede reducir solapamiento e ignorarlo puede aumentarlo. La evaluación humana y con GPT-5.2 Pro solo se muestra para LaMP-7. El paper no publica cuántos ejemplos fueron anotados, quiénes anotaron, acuerdo interanotador ni puntuaciones crudas. Las correlaciones se calculan sobre medias agregadas de cinco sistemas por cuatro estilos, como máximo 20 puntos, no a nivel de ejemplo, y carecen de incertidumbre. Por ello apoyan asociación direccional entre rankings, no que el juez sea casi indistinguible de humanos ni que mida persona verdadera. Los valores del backbone 8B en la Tabla 4 tampoco coinciden con la Tabla 2: por ejemplo, METEOR de LaMP-5 es 0,321 frente a 0,398.

La reproducibilidad es bloqueante. El enlace anónimo de código de v2 ha caducado. Un repositorio público homónimo contiene cinco scripts parciales para LaMP-7, pero no puede vincularse de forma inequívoca al archivo expirado y no incluye datos, resultados, checkpoints, dependencias, jueces, evaluación humana, tests o CI. El checkpoint guardado por entrenamiento usa claves anidadas incompatibles con las claves planas que exige inferencia; alpha no aparece como opción; la referencia DPO es otra política PsPLUG congelada desde inicialización, no el modelo base de las ecuaciones; y faltan las otras tareas y barridos de estilo. La contribución respalda que un perfil comprimido y un prefijo compartido son una vía plausible de personalización y que los estilos explícitos alteran los resultados de LaMP. No demuestra desentrelazado matemático, una compensación calibrada, eficiencia medida, privacidad, personalidad psicológica ni reproducción extremo a extremo. Bibliográficamente, v2 sigue siendo un preprint y reemplaza el título, abstract y orden de autores de v1.

Research question

Can a continuous prefix derived from history preserve user-specific patterns when the LLM receives an explicit style instruction, and does it allow scaling the trade-off between both signals at inference time?

Method

PsPLUG freezes Qwen3 and learns a prefix of three vectors: shared system, user profile, and query. The profile summarizes ten elements of the history with Qwen3-8B, is encoded with BGE, and is projected with an MLP. A Bradley-Terry/DPO loss prefers the user's historical response over a style-conditioned base generation. The evaluation compares six LaMP tasks without style, three generative tasks under four styles, five systems, three backbone sizes, five alpha intensities, task metrics, GPT-5.2 Pro judges, and an insufficiently documented human validation.

Sample: Six LaMP tasks with 6,542/1,500, 5,073/1,410, 20,000/2,500, 12,500/1,500, 14,682/1,500, and 13,437/1,498 train/validation instances, respectively. The mean histories range from 15.7 to 204.6 elements, but the profile uses only k=10. Unique users per split and the size of the human subset are not reported. The published correlation aggregates five systems by four styles, at most 20 points per dimension.

Findings

  • Without style, PsPLUG is better in 12 of 15 metrics, ties one, and loses two; it does not literally surpass all baselines on all metrics.
  • In the 24 ROUGE cells with style, PsPLUG is first or second in 22 when recalculating the numerical ranks.
  • The typographic marks for best and second contain at least six errors that do not always match the values.
  • ROUGE against a non-stylized reference cannot separate user preservation and style obedience.
  • The human/LLM evaluation of LaMP-7 shows aggregate correlation, but does not validate example-level agreement or the other two generative tasks.
  • The 8B results in Table 4 diverge from the Qwen3-8B results in Table 2 without explanation.
  • Figure 5 shows sensitivity to alpha, but does not publish numbers, uncertainty, or an independent curve of style compliance.
  • Efficiency is argued via asymptotic complexity; latency, memory, throughput, or real time are not measured.

Limitations

  • The positive is not controlled by style and the negative is, so the training contrast is confounded.
  • There is no theorem or factorial proof that separately identifies persona and style.
  • Repeated runs, result seeds, intervals, significance, or sensitivity to the sampling of ten histories are not reported.
  • The human protocol omits sample size, annotators, blinding, assignment, agreement, and raw scores.
  • Correlations are computed over system-style means and may hide example-level disagreement.
  • The proprietary judge has no reproducible API/configuration snapshot and receives a reference potentially incompatible with the arbitrary style.
  • The tables have erroneous ranking marks and contradictory 8B values.
  • Only four hand-written styles, LaMP in English, and a narrow set of backbones are studied.
  • The official artifact is expired and the public candidate is partial, incompatible, and does not reproduce tables.
  • Privacy claims are not accompanied by client implementation, attacks, deletion, consent, or formal guarantees.

What the study does not establish

  • That the objective learns a pure personalization residual after controlling for style.
  • That the positive responses obey each imposed style.
  • That persona and style are identifiable or causally separable factors.
  • That ROUGE against the original reference simultaneously measures both objectives without confounding.
  • That PsPLUG is better on all metrics or that small differences are statistically reliable.
  • That LLM judges are nearly indistinguishable from humans or measure true persona at the example level.
  • That alpha provides a calibrated trade-off between two separately measured objectives.
  • That latency, memory, throughput, or multi-user scalability are minimal in real measurements.
  • That the behavior generalizes to other languages, domains, histories, open styles, or models.
  • That the public code runs the full pipeline or reproduces any table.
  • That the compressed profile preserves the entire history, does not hallucinate, or protects privacy.
  • That psychological personality, trait stability, or an internal representation of persona is measured.
  • That v2 has been peer-reviewed or published outside of arXiv.

Traceability

Scope: Full text

Version: arXiv:2601.06362v2, 20 pages; TeX source, v1-v2 metadata drift, expired anonymous code artifact and public homonymous candidate commit ea133d4 also audited

Consulted source: https://arxiv.org/abs/2601.06362v2

Review: Codex 20-page arXiv-v2 visual, TeX-source, v1-v2 metadata, objective-identification, table-ranking, human-LLM aggregation, efficiency and expired/candidate-code audit, 2026-07-16

Approval: Codex fidelity pass, 2026-07-16

English translation: approved, 2026-07-18

Models evaluated

  • Qwen3-8B
  • Qwen3 backbone variants labeled 4B and 32B (exact checkpoints not reported)
  • BAAI/bge-base-en-v1.5
  • GPT-5.2 Pro

Instruments and metrics

  • Profile-Augmented Generation user profiles
  • Three-token continuous prefix
  • Bradley-Terry/DPO-style preference loss
  • Inference-time user-vector scaling alpha
  • Accuracy, F1, MAE and RMSE
  • ROUGE-1, ROUGE-L and METEOR
  • LLM persona and style judges
  • Human ratings
  • Spearman, Kendall and Pearson correlations

Data used

  • LaMP-1 Citation Identification
  • LaMP-2 Movie Tagging
  • LaMP-3 Product Rating
  • LaMP-4 News Headline Generation
  • LaMP-5 Scholarly Title Generation
  • LaMP-7 Tweet Paraphrasing

Evidence and location

  • Current version, architecture, objective, experiments, tables, appendices, and limitations: arXiv:2601.06362v2, 20 pages; TeX source archive
  • Derives from title, abstract, and author order: arXiv v1 and v2 metadata records; v2 API and PDF
  • Status of the official artifact and scope/defects of the public candidate: Anonymous GitHub PsPLUG-038C repository_expired response; hulameow/PsPLUG commit ea133d4
  • Objective confounding, ranking recalculation, human evaluation, 8B contradiction, efficiency, and reproducibility: reports/verification/article-265-psplug-version-objective-metric-human-evaluation-efficiency-and-artifact-audit.json