Persona-DB: Efficient Large Language Model Personalization for Response Prediction with Collaborative Data Refinement

Personas, identity, and agents2025ACL AnthologyApproved editorial review

Authors: Chenkai Sun, Ke Yang, Revanth Gangi Reddy, Yi R. Fung, Hou Pong Chan, Kevin Small, ChengXiang Zhai, Heng Ji

Keywords: Large Language Models, Persona, Personalization, Retrieval-Augmented Generation, User Modeling, Collaborative Filtering, Response Prediction, Privacy and Profiling

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

8
Authors
14
Findings
26
Limitations
10
Evidence

Editorial summary

English

This COLING 2025 paper proposes Persona-DB, a retrieval-based personalization system that transforms user profiles and histories into a hierarchy of History, Distilled Persona, Induced Persona, and Cache; it embeds these representations to match similar users, retrieves information from the current and neighboring users, and asks GPT-3.5-turbo-0613 to predict a response. Persona here is not psychometric personality: it consists of LLM-generated summaries and inferences about values, ideology, interests, roles, opinions, and behavioral patterns. The paper evaluates two tasks. RFPN contains 13.3k responses from 8.4k users to 3.8k Twitter headlines and predicts sentiment polarity and intensity. OpinionQA predicts individual Pew survey responses for Biomedical and Food Issues, 67 questions and 2,537 respondents, Global Attitudes, 104/2,596, and America in 2050-90/2,524, with 1,000 test cases per topic. At RFPN top-40, Persona-DB reports 47.67 Spearman, 47.88 Pearson, 62.66 Micro-F1, and 50.59 Macro-F1, versus 44.89/45.05/59.96/47.32 for IntSum and 42.80/43.09/59.58/48.80 for History Full. For the 100 sparsest nonempty-history users, 13.81 interactions on average, it obtains 56.75 Pearson versus 45.41 for IntSum, a difference of 11.34 points that the paper describes as 11%. For the 300 most frequent users it reaches 49.71 versus 46.58. Efficiency evidence narrows as capacity falls: at top-10, Persona-DB exceeds IntSum by only 0.13 Pearson points and 0.20 accuracy points. Table 3 compares methods at equal top-k values of 40, 30, and 10 and does not itself establish the introduction's ten-times-smaller maximal retrieval claim. It also does not measure total cost, because hierarchical construction requires prior LLM and embedding calls. Figure 3 places Persona-DB above baselines on OpinionQA, but provides bars without exact values, errors, or outputs. Human validation scores 50 extractions using three graduate students and reports a 3.9/5 mean, without inter-rater agreement, dispersion, or a sufficiently reported rubric. Code audit materially changes the main interpretation. get_user_profile2 excludes the target question from the current user's history, but UserRetriever.get_user_analysis with nb_hist_only=True includes every neighboring question-answer pair, the source comment explicitly says it includes the current query. join_db concatenates these records before retrieval into the prompt. In OpinionQA, where many respondents answer the same survey questions, collaborative context can therefore contain neighbors' answers to the exact target item. Without rerunning evaluation after excluding post_id from every neighboring database, the reported gain cannot distinguish persona generalization from same-item collaborative filtering and label leakage. History Full is not literally complete either: both runners remove records until histories fit 2,000-2,500 tokens. The release has no tests, CI, Persona-DB data, README-referenced data.txt, derived personas, embeddings, executed prompts, predictions, or results; RFPN requires reconstructing tweets from pointers through an old API workflow, and OpinionQA preparation code is absent. Source compiles syntactically but uses a legacy environment, broad dependencies, unbounded retries, and no repository-wide license. The paper supplies point estimates from one run at temperature 1 without intervals, tests, replications, or bootstrap. The defensible conclusion is narrow: hierarchical profile abstractions and collaborative retrieval may improve response prediction on these datasets, especially with larger retrieval capacity and sparse users, but the study does not validate personality, profile fidelity, unseen-user or unseen-question generalization, or operational efficiency. Potential target-question leakage prevents clean attribution of OpinionQA results to the persona mechanism.

Español

Este trabajo de COLING 2025 propone Persona-DB, un sistema de personalización por recuperación que transforma perfiles e historiales de usuario en una jerarquía de History, Distilled Persona, Induced Persona y Cache; busca usuarios similares con embeddings, recupera información del usuario y sus vecinos y pide a GPT-3.5-turbo-0613 que prediga una respuesta. Aquí «persona» no es personalidad psicométrica: son resúmenes e inferencias generados por LLM sobre valores, ideología, intereses, roles, opiniones y patrones de conducta. El paper evalúa dos tareas. RFPN contiene 13,3 mil respuestas de 8,4 mil usuarios a 3,8 mil titulares de Twitter y predice polaridad e intensidad de sentimiento. OpinionQA predice respuestas individuales a preguntas Pew en Biomedical and Food Issues, 67 preguntas, 2.537 personas, Global Attitudes, 104/2.596, y America in 2050, 90/2.524, con 1.000 casos de test por tema. En RFPN top-40, Persona-DB declara Spearman 47,67, Pearson 47,88, Micro-F1 62,66 y Macro-F1 50,59, frente a IntSum 44,89/45,05/59,96/47,32 y History Full 42,80/43,09/59,58/48,80. Para los 100 usuarios con historiales no vacíos más escasos, 13,81 interacciones de media, obtiene Pearson 56,75 frente a 45,41 de IntSum: +11,34 puntos, que el texto presenta como 11 %. Para los 300 usuarios más frecuentes logra 49,71 frente a 46,58. Pero la evidencia de eficiencia se estrecha al reducir capacidad: con top-10, Persona-DB solo supera a IntSum por 0,13 puntos Pearson y 0,20 de accuracy. Table 3 compara métodos con los mismos top-k 40/30/10 y no demuestra por sí sola la afirmación introductoria de un tamaño máximo diez veces menor. Tampoco informa coste total: la construcción jerárquica requiere llamadas y embeddings previos. Figure 3 muestra Persona-DB por encima de baselines en OpinionQA, pero solo como barras sin valores exactos, errores o outputs. La validación humana puntúa 50 extracciones con tres estudiantes y media 3,9/5, sin acuerdo entre evaluadores, dispersión o rúbrica suficiente. La auditoría del código cambia además la interpretación principal. get_user_profile2 elimina la pregunta objetivo del historial del usuario actual, pero UserRetriever.get_user_analysis con nb_hist_only=True incluye todos los pares pregunta-respuesta de los vecinos, el propio comentario dice que incluye la query actual. join_db los concatena y recupera elementos para el prompt. En OpinionQA, donde muchas personas contestan las mismas preguntas, el contexto colaborativo puede contener la respuesta de vecinos al ítem objetivo. Sin una reevaluación que excluya post_id de todas las bases vecinas, la mejora no separa generalización de persona, filtrado colaborativo de mismo ítem y fuga de etiqueta. El baseline History Full tampoco es literalmente completo: ambos runners eliminan registros hasta quedar en 2.000-2.500 tokens. No hay tests, CI, datos Persona-DB, data.txt nombrado por README, personas derivadas, embeddings, prompts ejecutados, predicciones ni resultados; RFPN exige reconstruir tweets desde punteros con una API antigua y falta el pipeline de preparación OpinionQA. El código compila sintácticamente, pero usa un entorno legado, dependencias amplias, retries indefinidos y carece de licencia general. El paper solo ofrece point estimates de un run, aunque usa temperatura 1; no publica intervalos, tests, repeticiones o bootstrap. La conclusión defendible es limitada: abstracciones jerárquicas y recuperación colaborativa pueden mejorar predicción de respuestas en estos datasets, especialmente con más capacidad y en usuarios escasos, pero el estudio no valida personalidad, fidelidad del perfil, generalización a usuarios o preguntas nuevas ni eficiencia operacional. La fuga potencial de la pregunta objetivo impide atribuir limpiamente los resultados OpinionQA al mecanismo de persona.

Research question

Can a hierarchical base of profiles and histories, enriched with data retrieved from similar users, maintain or improve the prediction of personal responses using fewer context elements than conventional histories, recency, or summaries?

Method

GPT-3.5-turbo-0613 distills the history into persona fields and then induces more general beliefs. A high-level category Cache is used to represent and match users with text-embedding-ada-002. For each query, top-k elements from the user and from neighbors are retrieved, concatenated in a fixed proportion, and GPT-3.5 predicts RFPN polarity/intensity or an OpinionQA option. The audit reconciles tables, inspects the 16 published pages, compiles 34 Python sources, and traces exclusion of post_id, JOIN, token limits, metrics, missing data, dependencies, and license in the official commit.

Sample: RFPN declares 13.3 thousand responses from 8.4 thousand users to 3.8 thousand headlines; its splits are 10,977/1,341/1,039 rows and 7,243/1,206/961 users. OpinionQA uses three Pew waves/topics with 2,524-2,596 respondents and 67-104 questions, evaluating 1,000 examples per topic. Cold-start cases are the 100 shortest non-empty histories, mean 13.81, and frequent cases are the 300 longest. The records necessary to verify overlap of users, questions, or leakage at scale are not published.

Findings

  • The authoritative source is COLING 2025, Anthology ID 2025.coling-main.20, pages 281-296; the 16 pages were rendered and visually inspected.
  • RFPN top-40: Persona-DB 47.67 Spearman, 47.88 Pearson, 62.66 Micro-F1, and 50.59 Macro-F1.
  • Against IntSum top-40 the differences are +2.78 Spearman, +2.83 Pearson, +2.70 Micro-F1, and +3.27 Macro-F1.
  • Cold-start: Pearson 56.75 against 45.41 for IntSum, +11.34 points on a 0-100 scale; not a universal relative improvement.
  • Frequent users: Persona-DB 49.71 Pearson against 46.58 for IntSum and 45.07 for History Full.
  • At top-10, the advantage over IntSum drops to 0.13 Pearson and 0.20 accuracy.
  • Table 3 only compares methods with equal top-k; it does not contain a comparison demonstrating ten times less maximum capacity.
  • OpinionQA is published as a chart without exact figures, uncertainty, or downloadable results.
  • Validation of 50 extractions obtains a mean of 3.9/5, but does not publish agreement, dispersion, sampling, or error analysis.
  • The code removes post_id only from the evaluated user's history and retains the same question in neighbor histories.
  • UserRetriever.get_user_analysis in nb_hist_only mode explicitly includes the neighbor's current query; JOIN delivers those pairs to retrieval.
  • History Full is truncated in code to a limit of 2,000-2,500 tokens.
  • The three train/dev/test OpinionQA objects load the same test file; one case path concatenates the three copies.
  • The 34 Python files compile, but the repository contains no data, outputs, tests, results, or general license.

Limitations

  • The target question is not excluded from all collaborative bases; potential answer leakage exists in OpinionQA.
  • A leakage-free ablation comparing JOIN with neighbors without the same post_id is not published.
  • A held-out split by user, question, or survey wave for OpinionQA is not documented.
  • Transfer to new users, questions, platforms, languages, or cultures is not tested.
  • Persona means LLM-generated profile inferences, not measured personality.
  • The prompts infer values, ideology, morality, roles, profession, possessions, and potentially sensitive interests.
  • The 3.9/5 validation lacks inter-rater agreement, CI, distribution, and category-level audit.
  • There is no evaluation of factual accuracy, temporal stability, or consent of the inferred individuals.
  • There is no fairness, subgroup, calibration, privacy, membership inference, or sensitive-attribute leakage evaluation.
  • Collaborative composition may share responses and sensitive profiles among users and amplify stereotypes.
  • The paper acknowledges risks of stereotype, surveillance, targeting, consent, and transparency, but the code implements no mitigations.
  • The results are point estimates from a single run; there are no tests, CI, repetitions, bootstrap, or public predictions.
  • Temperature 1 and API seed do not guarantee repeatability or estimate variance.
  • The top-10 benefit over IntSum is minimal and could be within unmeasured variation.
  • Efficiency is evaluated by number of elements, not by tokens, latency, dollars, storage, or total compute.
  • The prior cost of hierarchical extraction and embeddings is not quantified or amortized.
  • History Full does not necessarily represent the full history due to token truncation.
  • Figure 3 does not provide exact OpinionQA values or error bars.
  • RFPN is not distributed: reconstructing tweets from pointers and changed APIs may be impossible.
  • OpinionQA preparation, data.txt, derived personas, embeddings, prompts, outputs, and tables are missing.
  • The OpenAI models used are proprietary retired services; there are no request-response traces.
  • The Python 3.7/PyTorch 1.9 environment and numerous outdated dependencies make current execution difficult.
  • The API retries indefinitely without timeout, retry limit, cost cap, or structured logging.
  • Malformed OpinionQA outputs fall back to class 0, mixing generation error with substantive response.
  • No top-level license exists, and the vendored BLEU scorer retains a restrictive notice alongside an MIT license.
  • There is no prospective study of experience, utility, or harm in real users.

What the study does not establish

  • It does not demonstrate that Persona-DB measures personality or psychometric traits.
  • It does not validate that the generated profiles are true, stable, or consented.
  • It does not demonstrate that the OpinionQA improvement survives eliminating the target question from all neighbors.
  • It does not cleanly separate generalizable persona from collaborative filtering by same item or label leakage.
  • It does not establish generalization to new users or questions.
  • It does not demonstrate ten times fewer tokens, cost, latency, storage, or total compute.
  • It does not demonstrate statistical significance or robustness of small differences.
  • It does not demonstrate that History Full is a truly complete history baseline.
  • It does not establish privacy, fairness, or safety for sensitive profiling.
  • It does not reproduce the published tables from the official artifact.

Traceability

Scope: Full text

Version: Proceedings of COLING 2025, Anthology ID 2025.coling-main.20, pages 281-296, 16 pages

Consulted source: https://aclanthology.org/2025.coling-main.20/

Review: Codex complete bilingual full-text fidelity pass using the published COLING version, all-page visual inspection, official repository and linked RFPN artifact audit, exact Table 1-3 reconciliation, neighbor target-question leakage tracing, History Full truncation review, OpinionQA split-path audit, extraction-validation assessment, construct-boundary correction, efficiency-claim review, dependency and license inventory, and end-to-end reproducibility assessment; summaries written from paper and artifact evidence rather than abstract keywords, 2026-07-15

Approval: Codex fidelity pass, 2026-07-15

English translation: approved, 2026-07-18

Models evaluated

  • openai/gpt-3.5-turbo-0613 for persona extraction and downstream prediction
  • openai/text-embedding-ada-002 for user and item retrieval
  • microsoft/llmlingua-2-xlm-roberta-large-meetingbank comparison
  • TheBloke/Llama-2-7b-Chat-GPTQ optional legacy LLMLingua paths in released code

Instruments and metrics

  • LLM-generated Distilled Persona, Induced Persona, and Cache fields
  • Cosine-similarity user and item retrieval
  • Collaborative JOIN by concatenating neighbor database entries
  • RFPN sentiment polarity and ordinal intensity prediction
  • OpinionQA multiple-choice response prediction
  • Spearman and Pearson correlations
  • Micro-F1 and Macro-F1; Micro-F1 acts as accuracy for single-label classification
  • MSE and one-minus-normalized Wasserstein alignment score
  • Three-rater 50-item extraction-quality score
  • Independent target-leakage, construct-boundary, efficiency, artifact, and licensing audit

Data used

  • RFPN: reported 13.3k reactions, 8.4k users, and 3.8k Twitter news headlines
  • RFPN train 10,977; development 1,341; test 1,039
  • OpinionQA Biomedical and Food Issues: 67 questions and 2,537 respondents
  • OpinionQA Global Attitudes: 104 questions and 2,596 respondents
  • OpinionQA America in 2050: 90 questions and 2,524 respondents
  • OpinionQA reported 1,000-sample test split per topic
  • No Persona-DB dataset, derived persona, prediction, embedding, or result files released

Evidence and location

  • Scope, hierarchy, and JOIN mechanism: Published pages 281-284, Abstract and Sections 1-2
  • Datasets, models, metrics, and Table 1: Published pages 285-286, Sections 3.1-3.3 and Table 1
  • Cold-start, frequent users, and retrieval capacity: Published page 287, Sections 3.4-3.5 and Tables 2-3
  • Composition, conclusions, limitations, and ethics: Published pages 288-289, Sections 3.6-5, Limitations and Ethics Statements
  • Splits, human validation, Wasserstein, LLMLingua, and prompts: Published pages 292-296, Appendix A and Tables 4-6, Figures 5-13
  • Potential target question leakage in neighbors: Official repository commit c6bb783, utils/data_utils.py get_user_profile2/get_user_analysis/join_db and llm_predict_opinionqa.py audited 15 July 2026
  • History Full truncation and OpinionQA train/dev/test duplication: Official repository commit c6bb783, llm_predict.py and llm_predict_opinionqa.py audited 15 July 2026
  • Reproducibility, dependencies, missing files, and license: Official repository tree at commit c6bb783 and linked response_forecasting commit 808f738 audited 15 July 2026
  • Integral audit of leakage, construct, efficiency, and artifact: reports/verification/article-194-persona-db-leakage-and-artifact-audit.json
  • Complete visual inspection: All 16 published COLING PDF pages rendered and visually inspected on 15 July 2026