SynthesizeMe! Inducing Persona-Guided Prompts for Personalized Reward Models in LLMs

Trait induction and control2025ACL AnthologyApproved editorial review

Authors: Michael J. Ryan, Omar Shaikh, Aditri Bhagirath, Daniel Frees, William Held, Diyi Yang

Keywords: Personalized Reward Modeling, Synthetic User Personas, Pairwise Preferences, LLM-as-a-Judge, PersonalRewardBench

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

6
Authors
10
Findings
22
Limitations
6
Evidence

Editorial summary

English

SynthesizeMe turns a short history of pairwise choices into a personalized prompt for predicting future choices. It does not start from a declared identity or trait inventory. First, an LLM speculates about why the user selected one response and validation-driven search retains reasoning that helps predict a small held-out context. The selected explanations are then condensed into a natural-language persona. Finally, the method searches for prior examples that, together with the persona, maximize validation accuracy. The instruction for writing personas is optimized with DSPy MIPROv2 on PRISM users and transferred to Chatbot Arena and other models. The paper explicitly states that the initial reasoning is generated without background knowledge and is purely speculative.

Evaluation uses PersonalRewardBench, a curation of two existing datasets. Chatbot Arena contributes 131 users and 1,338 pairs; PRISM contributes 723 users, 3,897 conversations, and 16,705 pairs, for 854 users in total. Users need at least five preferences, GPT-4o-mini filters queries deemed personalizable, and five LLM judges run in both response orders remove unanimously judged cases. Users are partitioned 40/10/50% into train, validation, and test; within each user, temporally ordered interactions are divided 50/20/30% into training context, validation context, and target preferences. Full Parquet QA reproduces the totals and finds no user or conversation crossing partitions. It also finds 12 duplicate PRISM rows, six pairs with identical chosen and rejected responses, and up to 16 correlated pairs per conversation. The personalizability filter was checked on only 100 manually labeled examples with a 50-case test set, without annotator agreement or uncertainty.

For LLM judges, the paper reports gains up to 4.4 points on Chatbot Arena and 3.41 on PRISM. Those figures depend on the variant and comparator: 61.97% for 70B Just Demos versus the 57.57% Memory baseline on Arena, and 57.76% for 70B Just Demos versus the 54.35% default judge on PRISM. The ablation is central: persona text without demonstrations is worse than the default judge for 3B and 70B on both datasets and improves only the two 8B cases. Every winning configuration contains selected examples. With a persona prompt distilled from 70B, the 8B Persona+Demos variant reaches 61.62% on Arena and 55.24% on PRISM, versus 53.70% and 52.80% for the default judge. Distribution-fitted Bradley-Terry reward models are much stronger, however. Adding SynthesizeMe improves only three of six configurations slightly, always within the reported intervals, and degrades or leaves unchanged the others; the authors therefore recommend it mainly for in-context personalization.

Persona fidelity has a narrow operational meaning. PRISM users wrote one or two sentences about what they wanted from an LLM; GPT-4o-mini, not human raters, decides whether a generated persona strongly matches those statements. True-pair match rates rise from 26.5% to 50.2% and 56.1% for 3B, 8B, and 70B, while random matches remain around 46-47% for 8B/70B; only the 70B comparison is marked significant. This supports some predictive association, not the truth of biographical or psychological detail. Appendix prompts invent unobserved names, ages, occupations, values, interests, and traits. The optimized 70B instruction itself embeds topics from the optimization set, art, social justice, environmental issues, pessimism, and hope, showing dataset imprinting. Readable text is not automatically a causal explanation or factual user description.

Uncertainty and artifacts further narrow the conclusion. The released code bootstraps the two orders of a pair and multiple pairs from one conversation as separate observations, without user or conversation clustering, so uncertainty may be understated. The benchmark is selected with the same class of LLM judges later evaluated, and unanimity is not ground-truth accuracy. The authors acknowledge possible contamination of Gemini models by public Chatbot Arena data, limited longitudinal realism, and sycophancy risk. The repository contains the core algorithm and prompts but not the full filtering pipeline or scripts to reproduce tables, baselines, and seeds; it has no tests and only a publishing workflow. `pip install SynthesizeMe` currently selects stable 0.0.1 rather than repository version 0.0.11-alpha.1 and lacks the documented dataset loader. The repository version omits the `datasets` dependency and has defects in loading, demo fallback, and pre-fit state. Both complete derived datasets are public, but their cards are empty and unlicensed, fail to carry forward upstream no-reidentification and licensing terms, and retain PRISM identifiers linkable to demographic profiles. The supported result is that selected examples and, in some settings, persona hypotheses improve preference prediction on two filtered benchmarks. The paper does not establish that the system discovers who a user is or that its personas are true, safe, or stable.

Español

SynthesizeMe convierte un historial breve de elecciones entre dos respuestas en un prompt personalizado para predecir elecciones futuras. No parte de una identidad o de rasgos declarados. Primero pide a un LLM que especule por qué el usuario eligió una respuesta y descarte las explicaciones que no ayudan a predecir un pequeño conjunto de validación. Después resume las explicaciones seleccionadas como una persona en lenguaje natural. Por último busca qué ejemplos previos, junto con esa persona, maximizan la precisión de validación. El prompt que define cómo redactar personas se optimiza con DSPy MIPROv2 en usuarios de PRISM y se transfiere a Chatbot Arena y a otros modelos. El propio paper aclara que el razonamiento inicial se produce sin información de fondo y es especulativo.

La evaluación usa PersonalRewardBench, una curación de dos datasets existentes. Chatbot Arena aporta 131 usuarios y 1.338 pares; PRISM, 723 usuarios, 3.897 conversaciones y 16.705 pares, para un total de 854 usuarios. Se conservan usuarios con al menos cinco preferencias, GPT-4o-mini filtra consultas consideradas personalizables y cinco jueces LLM, ejecutados en ambos órdenes de respuesta, eliminan los casos de acuerdo unánime. Los usuarios se separan 40/10/50% para train, validación y test; dentro de cada usuario, las interacciones temporales se reparten 50/20/30% entre contexto de entrenamiento, validación y objetivo. La auditoría completa de los Parquet confirma los totales y que no hay usuarios ni conversaciones cruzando particiones. También encuentra 12 filas PRISM duplicadas, seis pares con respuestas elegida y rechazada idénticas y hasta 16 pares correlacionados por conversación. El filtro de personalizabilidad solo se comprobó con 100 ejemplos manuales y 50 casos de test, sin acuerdo entre anotadores ni intervalos.

En los jueces LLM, el paper informa mejoras máximas de 4,4 puntos en Chatbot Arena y 3,41 en PRISM. Esas cifras requieren leer qué variante y baseline se comparan: 61,97% para Llama 70B con solo demostraciones frente a 57,57% del baseline Memory en Arena, y 57,76% con solo demostraciones frente a 54,35% del juez por defecto en PRISM. La ablación es decisiva: la persona sin demostraciones empeora al juez por defecto en los modelos 3B y 70B de ambos datasets y solo mejora los dos casos 8B. Todas las configuraciones ganadoras incluyen ejemplos seleccionados. Con el prompt de persona destilado desde 70B, Persona+Demos llega a 61,62% en Arena y 55,24% en PRISM con Llama 8B, frente a 53,70% y 52,80% del juez por defecto. Sin embargo, reward models Bradley-Terry ajustados a cada distribución son bastante más fuertes. Añadir SynthesizeMe solo mejora ligeramente tres de seis configuraciones, siempre dentro de los intervalos, y empeora o no cambia el resto; por eso los autores lo recomiendan principalmente para personalización in-context.

La supuesta fidelidad de las personas tiene una frontera estrecha. Los usuarios de PRISM escribieron una o dos frases sobre lo que querían de un LLM; GPT-4o-mini, no personas evaluadoras, decide si una persona generada coincide fuertemente con esas frases. Las coincidencias verdaderas pasan de 26,5% a 50,2% y 56,1% para 3B, 8B y 70B, mientras los emparejamientos aleatorios rondan 46-47% en 8B/70B; solo 70B está marcado como significativo. Esto valida cierta asociación predictiva, no la verdad de los detalles biográficos o psicológicos. Los anexos muestran prompts que inventan nombre, edad, profesión, valores, intereses y rasgos no observados. El prompt 70B optimizado incluso incorpora temas concretos del conjunto de optimización, arte, justicia social, medio ambiente, pesimismo y esperanza, evidencia de impronta del dataset. Que el texto sea legible no lo convierte en una explicación causal ni en una descripción factual del usuario.

La incertidumbre y los artefactos también limitan la conclusión. El código publicado calcula intervalos re-muestreando por separado las dos órdenes del mismo par y los múltiples pares de una conversación, sin clustering por conversación o usuario, por lo que puede subestimar incertidumbre. El benchmark selecciona casos mediante los mismos tipos de jueces LLM que después se evalúan, y acuerdo unánime no equivale a corrección. Los autores reconocen posible contaminación de modelos Gemini con Chatbot Arena público, bajo realismo longitudinal y riesgo de sicofancia. El repositorio contiene el algoritmo central y los prompts, pero no la construcción completa del benchmark ni scripts para reproducir tablas, baselines o seeds; no tiene tests y su único workflow publica paquetes. `pip install SynthesizeMe` instala hoy la versión estable 0.0.1, que no coincide con el repositorio 0.0.11-alpha.1 ni expone el loader documentado. La versión del repositorio omite la dependencia `datasets` y presenta fallos en carga, fallback de demos y estado previo al entrenamiento. Los dos datasets derivados son públicos y completos, pero sus tarjetas están vacías y sin licencia, no trasladan las condiciones de no reidentificación y licencias de las fuentes, y conservan identificadores PRISM enlazables con perfiles demográficos. La evidencia sólida es que seleccionar ejemplos y, a veces, añadir una hipótesis de persona puede mejorar predicción de preferencias en dos benchmarks filtrados. No demuestra que el sistema descubra quién es el usuario ni que las personas sean verdaderas, seguras o estables.

Research question

Can a system induce a textual persona and select demonstrations from 5-15 paired preferences to improve prediction of future preferences for each user without relying on predefined demographic categories?

Method

Speculative explanations are generated for training preferences and selected by accuracy on validation preferences. A LLM summarizes the explanations into a persona using a prompt optimized with MIPROv2, and a second search chooses informative demonstrations. Persona and examples condition a LLM-as-a-judge or a reward model. The evaluation compares ablations, memory and demography baselines, GPO/VPL/PAL, Bradley-Terry LoRA and transfer between model families on PersonalRewardBench.

Sample: PersonalRewardBench contains 131 users from Chatbot Arena, 1,338 conversations and 1,338 pairs, plus 723 users from PRISM, 3,897 conversations and 16,705 pairs: 854 users in total. The user splits are 23/19/89 and 280/65/378. Each user contributes at least five pairs and usually has fewer than 25; within user, 50/20/30% is used for train context, validation and target. The personalization filter was developed with 100 manual labels. The artifacts reproduce these totals, except for the typo of 720 PRISM users in Table 7.

Findings

  • The maximum reported improvements are 4.4 points on Chatbot Arena and 3.41 on PRISM, both achieved by the Just Demos variant with Llama 70B against different comparators.
  • The persona without demonstrations falls below the default judge on 3B and 70B on both datasets; it only improves the 8B conditions.
  • All winning configurations across the six main experiments include selected demonstrations.
  • Persona+Demos with distilled 70B prompt reaches 61.62% on Arena and 55.24% on PRISM with Llama 8B, compared to 53.70% and 52.80% for the default baseline.
  • Bradley-Terry reward models adjusted to the distribution widely outperform in-context judges.
  • Adding SynthesizeMe to the reward model only improves three of six cells, always within intervals, and worsens or ties the rest.
  • SynthesizeMe outperforms the default judge in 12 of 14 transfer tests, but only three of six repeated Qwen conditions reach the permutation thresholds.
  • Automatic persona-preference agreement reaches 56.1% on 70B versus approximately 47% random and only that scale is marked p<0.05.
  • The public data reproduce 854 users and the splits without leakage between users or conversations.
  • The repository and the two datasets exist, but do not allow end-to-end reproduction of the paper tables and the default installable package is misaligned.

Limitations

  • The initial reasoning about the user is explicitly speculative and is post-selected with very little data.
  • Persona, demos and global prompt are chosen by validation, with risk of overfitting and winner's curse without nested evaluation.
  • The persona alone does not improve consistently; the main benefit may come from the demonstrations.
  • Fidelity is judged with GPT-4o-mini, not with human annotation, and only against 1-2 sentences of stated preferences.
  • The optimized prompts invent biography and psychographics and incorporate topics from the optimization set.
  • The personalization filter only has 100 labels and lacks agreement, balance, intervals and error analysis.
  • The quality filter selects LLM judge disagreement; unanimity does not equate to correctness.
  • Benchmark selection and evaluation depend on LLM-as-a-judge, creating a model-mediated circuit.
  • PRISM contains multiple correlated pairs per conversation and six pairs of identical responses.
  • The code intervals do not cluster by user or conversation and duplicate each pair in two orders.
  • Temporal change, topic preference or interaction is not modeled; the user is treated as a stable function.
  • Users with fewer than five preferences are excluded.
  • The evaluation is concentrated in English and on selected subjective/controversial cases.
  • GPT and Gemini results are single-run and there may be contamination from public Chatbot Arena.
  • LLoom clusters describe generated texts, without human validation or stability.
  • The repository does not include full pipeline, table scripts, intermediate data, seeds or environment lock.
  • There are no tests; Ruff detects 22 incidences under the declared configuration.
  • PyPI installs an obsolete stable version and the repo version omits the datasets dependency.
  • load(), fallback without demos, output directory and pre-fit guards present functional defects.
  • The derivative cards are empty, without license, provenance, PII or withdrawal process.
  • PRISM IDs allow linking conversations with demographic profiles and stated preferences.
  • Personalization may amplify stereotypes, manipulation and sycophancy; mitigations are not evaluated.

What the study does not establish

  • That the generated persona is a factually or psychologically valid description of the user.
  • That the system discovers identity, age, profession, values, interests or real traits.
  • That the textual persona, separated from the demonstrations, causes the main improvements.
  • That agreement decided by GPT-4o-mini equates to human validation.
  • That PersonalRewardBench represents ordinary unfiltered traffic.
  • That unanimous agreement of five judges means 100% accuracy.
  • That the published intervals have correct coverage under dependence by pair, conversation and user.
  • That the results generalize to new users with little data, other languages or longitudinal preferences.
  • That transfer to closed models is free of contamination.
  • That avoiding demography as input eliminates the risk of identity linkage or sensitive profiling.
  • That the personas are safe against stereotypes, sycophancy or manipulation.
  • That the current package works cleanly following the README.
  • That the public artifacts reproduce all the experiments and tables of the article.
  • That the derived datasets correctly carry over licenses, provider terms and re-identification prohibitions.

Traceability

Scope: Full text

Version: ACL 2025 long paper; audited text arXiv:2506.05598v1, 34 pages; official code, PyPI package and both full PersonalRewardBench derivatives also audited

Consulted source: https://aclanthology.org/2025.acl-long.397/

Review: Codex 34-page full-text visual, ACL publication, full benchmark Parquet, persona-validity, selection/statistics, upstream governance, GitHub/PyPI/package and claim-boundary audit, 2026-07-16

Approval: Codex fidelity pass, 2026-07-16

English translation: approved, 2026-07-18

Models evaluated

  • Llama-3.2-3B-Instruct
  • Llama-3.1-8B-Instruct
  • Llama-3.3-70B-Instruct
  • Llama-3.1-70B-Instruct prompt optimizer
  • GPT-4o-mini
  • Gemini-2.0-Flash, Gemini-2.5-Flash and Gemini-2.5-Pro
  • Qwen3-8B, Qwen3-30B-A3B and Qwen3-32B
  • GPT-4o-mini, Llama-3.1-70B, Llama-3.3-70B, Gemini-1.5-Pro and Qwen2.5-72B quality-filter judges

Instruments and metrics

  • Pairwise preference prediction accuracy
  • DSPy MIPROv2 persona-prompt optimization
  • Bootstrap reasoning with validation selection
  • Validation-selected few-shot demonstrations
  • Default, demographic and memory LLM-as-a-judge prompts
  • Bradley-Terry rank-64 LoRA reward models
  • GPO, VPL and PAL personalized reward-model baselines
  • GPT-4o-mini personalizability filter
  • Five-model bidirectional consensus filter
  • GPT-4o-mini persona-to-stated-preference match test
  • LLoom persona clustering
  • Item-level bootstrap confidence intervals and permutation tests

Data used

  • PersonalRewardBench Chatbot Arena derivative: MichaelR207/chatbot_arena_personalized_0125
  • PersonalRewardBench PRISM derivative: MichaelR207/prism_personalized_0125
  • Upstream Chatbot Arena conversations
  • Upstream PRISM Alignment survey and conversations

Evidence and location

  • Publication, authorship, DOI, venue, pages and license: ACL Anthology 2025.acl-long.397 checked 2026-07-16
  • Method, tables, ablations, optimized prompts, ethics and limitations: arXiv:2506.05598v1 PDF, 34 pages; every page rendered and visually inspected
  • Totals, splits, duplicates, dependence by conversation, identical responses and possible direct identifiers: Full Parquet audit of MichaelR207/chatbot_arena_personalized_0125@636706aac659eb8da30c8e687ea22208eddc830f and MichaelR207/prism_personalized_0125@a03cab5122c397f1307ed092423a05732cc2aa93
  • Code, tests, wheel, lint, dependencies, PyPI and runtime defects: SALT-NLP/SynthesizeMe commit e91f340122db42b2940c07bcfe6a264bbe554353; clean installs, wheel build, compileall and Ruff audit on 2026-07-16
  • Licenses, terms, user_id linkage and upstream data governance: Official Chatbot Arena gated dataset terms and HannahRoseKirk/prism-alignment dataset card checked 2026-07-16
  • Persona validity, selection, uncertainty, reproducibility and claim limits: reports/verification/article-271-acl-synthesizeme-personalrewardbench-user-split-selection-persona-validity-code-data-and-claim-audit.json