$Ψ$-Bench: Evaluating Persona-Sensitive Influencing in Persuasive Dialogues

Applications, bias, and safety2026arXivApproved editorial review

Authors: Peixuan Han, Hongyi Du, Jiayu Liu, Yihang Sun, Yutong Liu, Jiaxuan You

Keywords: Personalized persuasion benchmark, Synthetic client simulation, Profile inference, LLM-as-judge evaluation, User modeling risks

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

6
Authors
8
Findings
13
Limitations
4
Evidence

Editorial summary

English

Ψ-Bench evaluates personalized persuasion through conversations between two models: the tested LLM attempts to influence a client instantiated by DeepSeek-v3.2 from a hidden profile. It contains three tasks. Viewpoint Debate uses the first 500 of 2,131 Change My View threads; Psychological Consultation uses 90 CounselBench questions; Everyday Request contains 100 GPT-4o-generated requests. Each dialogue lasts three rounds. Another DeepSeek-v3.2, which sees the profile and full conversation, assigns 1-to-9 scores for quality, personalization, and effect: opinion change, psychological improvement, or request acceptance. Debate profiles are written by DeepSeek from topic frequencies and LIWC attributes associated with the Reddit user. In the other tasks, PersonaMem-v2 profiles are sampled and refined to fit the query. These are not verified demographics or psychometric assessments. The prompts require every field to be filled, instruct the model to select one plausible value when evidence is absent, and prohibit uncertainty language; the so-called ground-truth profile therefore contains guesses about age, gender, origin, occupation, religion, family, politics, and personality. The client also receives a shared engineered resistance: firm views in Debate, severe distress and resistance to change in Consultation, or busyness and independence in Request. Ten models are tested. Quality usually exceeds 7/9, while mean Effect ranges from 4.05 for Qwen3-8B to 5.79 for GPT-5.1. GPT-5.1 leads Debate (6.12), Request (5.88), and the average; Qwen3-80B-A3B leads Consultation (6.10). Quality and Personalize correlate .752 and .772 with Effect, although all three scores come from the same judge over the same synthetic dialogues. On human CMV dialogues, judge Effect reaches .960 AUC for detecting delta; on CounselBench, Quality reaches .780 against the expert label, but no patient follow-up exists to validate Treatment Effect. When reconstructing the final human CMV response, the profiled client reaches .669 AUC versus .605 without a profile: this is a modest gain and does not compare generated language with that person's actual response. The human study asks five participants for profiles and creates 50 paired conversations totaling 150 turns. Judge means are similar for human and simulated clients, but correlations are only .45-.50; per-response fidelity and blind human preference are not measured. Oracle gives the persuader the same synthetic profile already conditioning the client and visible to the judge. The paper then reports +41.19% Personalize and +18.24% Effect. This demonstrates that rubric-aligned shared information helps inside the loop, not an 18.24% increase in persuasion of real people. The profile analyzer predicts a JSON profile from dialogue. Qwen3-4B-RL, trained with GRPO and BGE-M3 similarity to synthetic profiles, reports 55% versus 51.30% for DeepSeek-v3.2. However, the released CMV split is not user-independent: 102 identifiers occur in both train and test, and 149 of the 500 test queries belong to a persona already present in train. The analyzer sometimes exceeding Oracle with only about 50% profile similarity also suggests it may produce text useful for steering the persuader or satisfying the judge rather than recovering real facts. The official repository publishes queries, profiles, prompts, scripts, data releases, and a PyPI wheel, but not generations, item scores, the human study, LlamaGuard decisions, GRPO training, or Qwen3-4B-RL weights. It also has failures that prevent treating it as a reference implementation: the recommended CLI does not pass --task, so counsel and request execute with CMV prompts and metrics; generation mutates every profile in memory and appends resistance again each turn; the source build fails because a nonexistent fig package is declared; vllm is imported but undeclared; and requirements contains an invalid character. There are no tests or CI. A further unaddressed privacy risk is that 1,636 inferred profiles, including age, gender, religion, and politics, are published under Reddit usernames alongside 2,131 direct thread URLs. The defensible contribution is a public dataset and experimental structure for comparing models inside a controlled personalized simulation. It does not establish realistic users, true profiling, clinical effectiveness, human behavior change, persuasive safety, or end-to-end reproducibility.

Español

Ψ-Bench evalúa persuasión personalizada mediante conversaciones entre dos modelos: el LLM probado intenta influir y DeepSeek-v3.2 interpreta a un cliente condicionado por un perfil oculto. Incluye tres tareas. Viewpoint Debate usa las primeras 500 de 2.131 discusiones Change My View; Psychological Consultation usa 90 preguntas de CounselBench; Everyday Request contiene 100 peticiones generadas con GPT-4o. Cada diálogo dura tres rondas. Otro DeepSeek-v3.2, que ve el perfil y la conversación completa, asigna de 1 a 9 a calidad, personalización y efecto: cambio de opinión, mejora psicológica o aceptación de la petición. Los perfiles de Debate los redacta DeepSeek a partir de frecuencias temáticas y rasgos LIWC asociados al usuario de Reddit. En las otras tareas se muestrean perfiles de PersonaMem-v2 y se refinan para encajar con la consulta. No son datos demográficos comprobados ni evaluaciones psicométricas. Los prompts exigen rellenar todos los campos, seleccionar un valor plausible cuando falta evidencia y expresarlo sin incertidumbre; por tanto, el llamado perfil ground truth incluye conjeturas sobre edad, género, origen, profesión, religión, familia, política y personalidad. El cliente recibe además una resistencia artificial común: firmeza en Debate, grave malestar y rechazo al cambio en Consultation, o vida ocupada e independencia en Request. Se prueban diez modelos. La calidad suele superar 7/9, pero el Effect medio va de 4,05 para Qwen3-8B a 5,79 para GPT-5.1. GPT-5.1 lidera Debate (6,12), Request (5,88) y el promedio; Qwen3-80B-A3B lidera Consultation (6,10). Quality y Personalize correlacionan 0,752 y 0,772 con Effect, aunque las tres cifras proceden del mismo juez sobre los mismos diálogos sintéticos. En diálogos humanos CMV, el Effect del juez alcanza AUC 0,960 para detectar delta; en CounselBench, Quality obtiene 0,780 frente a la etiqueta experta, pero no existe resultado posterior del paciente para validar Treatment Effect. Al reconstruir la última respuesta humana CMV, el cliente con perfil alcanza AUC 0,669 frente a 0,605 sin perfil: es una mejora modesta y no compara el texto generado con la respuesta de esa persona. El estudio humano pide a cinco participantes un perfil propio y genera 50 conversaciones emparejadas, 150 turnos. Las medias del juez son parecidas para cliente humano y simulado, pero las correlaciones solo son 0,45–0,50; no se mide fidelidad de cada respuesta ni preferencia humana ciega. La condición Oracle entrega al persuasor el mismo perfil sintético que ya condiciona al cliente y que ve el juez. El paper informa entonces +41,19% en Personalize y +18,24% en Effect. Esto demuestra que compartir información alineada con la rúbrica ayuda dentro del circuito, no que aumente un 18,24% la persuasión de personas reales. El analizador de perfiles intenta inferir un JSON desde el diálogo. Qwen3-4B-RL, entrenado con GRPO y recompensa de similitud BGE-M3 contra los perfiles sintéticos, informa 55% frente a 51,30% de DeepSeek-v3.2. Sin embargo, el split CMV publicado no es independiente por usuario: 102 identificadores aparecen en train y test, y 149 de las 500 consultas de test corresponden a una persona ya presente en train. Además, que el analizador supere a veces a Oracle con apenas 50% de similitud sugiere que puede producir texto útil para dirigir al persuasor o satisfacer al juez, no recuperar hechos reales. El repositorio oficial publica consultas, perfiles, prompts, scripts, release de datos y wheel PyPI, pero no salidas, puntuaciones por ítem, estudio humano, decisiones LlamaGuard, entrenamiento GRPO ni pesos Qwen3-4B-RL. Presenta fallos que impiden tratarlo como implementación de referencia: el CLI recomendado no pasa --task, de modo que counsel y request se ejecutan con prompts y métricas CMV; el generador modifica cada perfil en memoria y vuelve a añadir la resistencia en cada turno; el source build falla por un paquete fig inexistente; vllm se importa pero no se declara; y requirements contiene un carácter inválido. Tampoco hay tests o CI. Hay además un riesgo de privacidad no discutido: se publican 1.636 perfiles inferidos, con edad, género, religión y política, ligados a nombres de usuario y 2.131 URLs directas de Reddit. La contribución defendible es un conjunto público y una estructura experimental para comparar modelos dentro de una simulación personalizada controlada. No demuestra usuarios realistas, perfilado verdadero, eficacia clínica, cambio de conducta humano, seguridad persuasiva ni reproducibilidad end-to-end.

Research question

To what extent can different LLMs influence simulated clients conditioned by profiles through dialogue, does improving access to or inference of those profiles increase the scored effect, and does this simulation serve as an approximation to human users?

Method

Generative benchmark in three scenarios with client and judge DeepSeek-v3.2. 690 test queries are paired with synthesized profiles, three rounds are generated per model, and a judge that knows the profile scores quality, personalization, and effect. Judge and client are partially validated against CMV delta, CounselBench labels, and a five-person study. Hiding or revealing the profile is compared, and a Qwen3-4B is trained with GRPO to reconstruct profiles from conversation prefixes using BGE-M3 similarity as reward. The additional audit reproduces counts and split, inspects all prompts, the dataset, GitHub, release, PyPI, and the executable code.

Sample: The main benchmark uses 690 queries and three rounds per model. Debate has 500 test queries; Consultation 90; Request 100. The human study includes five persons, 50 conversations, and 150 turns. Profile validation uses a CMV split that shares 102 identifiers between train and test, affecting 149/500 test queries.

Findings

  • Within the client and judge DeepSeek-v3.2, GPT-5.1 obtains the highest mean Effect, 5.79/9.
  • Quality usually exceeds 7, while Effect remains approximately between 4 and 6.
  • Personalize correlates 0.772 with Effect, slightly more than Quality, but both are measures from the same judge.
  • Revealing the synthetic profile to the persuader produces +41.19% Personalize and +18.24% reported Effect.
  • The client with profile improves CMV reconstruction AUC from 0.605 to 0.669, without establishing individual fidelity.
  • Five humans and their simulations produce similar means, but only correlations 0.45-0.50.
  • Qwen3-4B-RL reaches 55% similarity with synthetic profiles and improves Effect within the benchmark.
  • The repository publishes partial data and code, but its main CLI runs two tasks with the wrong type.

Limitations

  • The profiles called ground truth are generated or refined by LLMs and not verified by the persons represented.
  • The prompts force inventing absent fields with certainty and prohibit expressing uncertainty.
  • Client and judge are DeepSeek-v3.2 and share profile, creating family, construct, and rubric circularity.
  • Effect measures reaction generated by an LLM, not human attitude, behavior, or therapeutic outcome.
  • Consultation simulates severe distress and therapy without clinical outcome, crisis protocol, or harm assessment.
  • The human study has five persons, insufficiently described methods, and does not measure response agreement.
  • The CMV split shares 102 persons and 149 test queries with identities seen in train.
  • Temperature one and a single run per cell leave rankings and percentages without uncertainty.
  • The safety analysis does not validate manipulation, pressure, dependence, stereotypes, or privacy.
  • Sensitive profiles are published linked to usernames and Reddit URLs.
  • Outputs, per-item scores, human study, LlamaGuard filters, RL training, or weights are not published.
  • The CLI loses --task, profiles are mutated each turn, the source build fails, and dependencies are missing.
  • There are no tests, CI, lockfile, root license, or complete traceability of APIs and costs.

What the study does not establish

  • It does not demonstrate an 18.24% increase in persuasion of real persons.
  • It does not demonstrate that the synthetic client reproduces decisions or language of a specific individual.
  • It does not demonstrate that inferred profiles are true, calibrated, or psychometrically valid.
  • It does not demonstrate efficacy, safety, or clinical adequacy in psychotherapy.
  • It does not demonstrate broad human validation from five participants and moderate correlations.
  • It does not demonstrate generalization of the analyzer to unseen users due to identity overlap.
  • It does not demonstrate absence of manipulation or safe transfer to high-risk domains.
  • It does not allow end-to-end reproduction of results with the current public artifact.

Traceability

Scope: Full text

Version: arXiv:2606.02754v1

Consulted source: https://arxiv.org/abs/2606.02754v1

Review: Codex twenty-nine-page full-text visual, TeX, prompt, dataset, split, synthetic-client, judge, human-study, privacy, safety, GitHub, PyPI and repository-code audit, 2026-07-17

Approval: Codex fidelity pass, 2026-07-17

English translation: approved, 2026-07-18

Models evaluated

  • Qwen3-8B
  • Qwen3-32B
  • Qwen3-80B-A3B
  • DeepSeek-v3.2
  • DeepSeek-v4-pro
  • Grok-4-fast
  • Gemini-3-flash
  • Gemini-3.1-pro
  • GPT-5-mini
  • GPT-5.1
  • Qwen3-4B
  • Qwen3-4B-RL
  • GPT-4o
  • LlamaGuard
  • BGE-M3

Instruments and metrics

  • DeepSeek-v3.2 LLM judge: Quality 1-9
  • DeepSeek-v3.2 LLM judge: Personalize 1-9
  • DeepSeek-v3.2 LLM judge: Effect 1-9
  • CMV Delta binary outcome
  • CounselBench expert score
  • LIWC features
  • BGE-M3 field-level profile similarity
  • Five-participant paired human/simulated-client study
  • Oracle and irrelevant-profile ablations

Data used

  • Webis-CMV-20 / released CMV subset: 2,131 threads, 1,631 train and 500 test
  • CounselBench subset: 90 questions and 360 expert-scored answers
  • Everyday Request: 100 GPT-4o-generated requests
  • PersonaMem-v2 profiles
  • Released persona records: 1,636 CMV, 84 counsel and 100 request
  • GRPO profile-analyzer training set: 6,400 reported examples

Evidence and location

  • Metadata and version: Official arXiv record and API entry 2606.02754v1, checked 2026-07-17
  • Method, results, prompts, human study, safety, and limitations: arXiv v1, all twenty-nine PDF pages and complete TeX source
  • Data, release, PyPI, and implementation status: Hanpx20/Psi-Bench commit 48f3aa1, GitHub release v0.1.0 and PyPI psi-bench 0.1.0 checked 2026-07-17
  • Audit of synthetic profiles, split, circularity, privacy, safety, code, and reproducibility: reports/verification/article-307-psi-bench-synthetic-ground-truth-profile-leakage-client-judge-circularity-privacy-safety-and-repository-audit.json