Synthetic Personality Review

Enhancing Persona Following at Decoding Time via Dynamic Importance Estimation for Role-Playing Agents

Trait induction and control2026arXivApproved editorial review

Authors: Yuxin Liu, Mingye Zhu, Siyuan Liu, Bo Hu, Lei Zhang

Keywords: Personality, Persona conditioning, Role-playing agents, Human simulation

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

5
Authors
12
Findings
28
Limitations
5
Evidence

Editorial summary

English

Accepted as an ICLR 2026 Poster, the paper proposes Persona Dynamic Decoding (PDD), a training-free method intended to strengthen, during decoding, the character-profile attributes that appear most relevant to the current context. Its PIE module first generates a response with the full profile and scores each attribute through the probability ratio of that same response with and without the attribute. PIA then computes probability-ratio rewards for the highest-ranked attributes, combines them using PIE weights and aims to reweight the next-token distribution under a KL penalty. The main configuration keeps two attributes, uses beta=1 and requires roughly three model distributions per token in addition to the initial importance computation.

Experiments use Qwen2.5-7B-Instruct and LLaMA-3-8B-Instruct on CharacterEval, BeyondDialogue and PERSONALITYBENCH, with additional Qwen2.5 3B and 14B tests. GPT-4o performs pairwise comparisons; CharacterRM scores five role-playing dimensions; and several LLMs plus five researchers judge five author-defined importance criteria. PDD has the highest open-model CharacterRM average: 2.85 versus 2.83 for the strongest baseline on Qwen and 2.81 versus 2.75 on LLaMA. It averages 4.57 on PERSONALITYBENCH for both models. It does not, however, lead every dimension as the text claims: another baseline is higher on four of five Qwen traits and on LLaMA extraversion and neuroticism. Pairwise comparisons usually show more wins than losses, but several win rates remain below 50%. In rebuttal, dynamic prompting scores 2.76 versus 2.85 for PDD; one reviewer maintained the rating because the 0.09 gain did not clearly justify the extra compute.

The defensible contribution is a contextual steering strategy that can increase explicit profile expression under the selected evaluators. Its theoretical and application claims need substantial qualification. Replacing an entropy with the log probability of one self-generated response does not establish a conditional-mutual-information estimator, and the proxy is circular: a response already conditioned on the profile determines which profile components are reinforced. More seriously, the published policy equations write the reward as a function of the history rather than the candidate token; as written, the exponential factor cancels and the policy remains the base model. First-token behavior, a zero reward norm and negative importance scores are also undefined, and no code exists to resolve the ambiguity. Statistical details, human-evaluation reliability, model snapshots, processed data and outputs are missing. Evidence is primarily from fictional characters and explicit Big Five prompts rather than social outcomes or human populations. One highlighted case even rewards a 13-year-old murderer persona for proposing to make people disappear without a trace, showing that stronger adherence can also amplify unsafe behavior.

Español

El artículo, aceptado como póster en ICLR 2026, propone Persona Dynamic Decoding (PDD), un método sin ajuste fino para reforzar durante la decodificación los atributos de una ficha de personaje que parecen más relevantes para el contexto. Su módulo PIE genera primero una respuesta con el perfil completo y asigna importancia a cada atributo mediante el cociente entre la probabilidad de esa misma respuesta con y sin el atributo. PIA calcula después recompensas de razón de probabilidades para los atributos más importantes, las combina con esos pesos y pretende reponderar la distribución del siguiente token bajo una penalización KL. En la configuración principal se conservan dos atributos, se usa beta=1 y se necesitan aproximadamente tres distribuciones del modelo por token, además del cálculo inicial de importancia.

Los experimentos emplean Qwen2.5-7B-Instruct y LLaMA-3-8B-Instruct en CharacterEval, BeyondDialogue y PERSONALITYBENCH, con pruebas adicionales en Qwen2.5 de 3B y 14B. GPT-4o realiza comparaciones por pares; CharacterRM puntúa cinco dimensiones de rol y varios LLM, junto con cinco investigadores, juzgan cinco criterios de importancia definidos por los autores. PDD obtiene el mejor promedio abierto de CharacterRM: 2,85 frente a 2,83 para el mejor baseline en Qwen y 2,81 frente a 2,75 en LLaMA. En PERSONALITYBENCH alcanza 4,57 de media con ambos modelos. No gana, sin embargo, en todas las dimensiones que el texto afirma: queda por debajo de algún baseline en cuatro de cinco rasgos con Qwen y en extraversión y neuroticismo con LLaMA. Las comparaciones por pares suelen arrojar más victorias que derrotas, pero varias tasas de victoria quedan por debajo del 50 %. En el rebuttal, dynamic prompting logra 2,76 frente a 2,85 de PDD; un revisor mantuvo su nota porque la mejora de 0,09 no compensaba claramente el mayor coste.

La contribución defendible es una estrategia de steering contextual que, según los evaluadores elegidos, puede aumentar la expresión explícita de perfiles. Las afirmaciones teóricas y de aplicabilidad requieren mucha cautela. Sustituir una entropía por la log-probabilidad de una única respuesta autogenerada no demuestra una estimación de información mutua condicional, y el proxy es circular: la respuesta ya condicionada por el perfil decide qué partes del perfil se reforzarán. Además, las ecuaciones publicadas de la política escriben la recompensa como función del historial, no del token candidato; tal como están formuladas, el factor exponencial se cancela y la política queda igual a la base. Tampoco se define el primer token, la norma cero ni el tratamiento de importancias negativas. No existe código que permita resolver la ambigüedad. Faltan detalles estadísticos, fiabilidad de la evaluación humana, snapshots de modelos, datos procesados y salidas. La validación se limita sobre todo a personajes ficticios y rasgos Big Five explícitos, no a resultados sociales o poblaciones humanas. Un caso incluso premia como mayor fidelidad que el perfil de un menor homicida proponga hacer desaparecer personas sin dejar rastro, mostrando que más adherencia también puede amplificar conducta insegura.

Research question

Can the attributes of a person that matter in each context be dynamically estimated using the model's own probabilities, and can that estimation be used to improve, without additional training, the fidelity of role-playing agents during decoding?

Method

PDD has two modules. PIE generates a response G with the full prompt T and computes for each attribute w_i the logarithm of Pr(G|T) divided by Pr(G|T_i), where T_i removes that attribute. PIA normally retains the two attributes with the highest score, computes for each a local likelihood-ratio reward over a two-token window, combines the rewards using the PIE weights, normalizes by their L2 norm, and applies an exponential solution of a KL-regularized objective. It is compared with simple and persona prompting, ICL, OPAD, and, on Big Five, NPTI and PAS. GPT-4o judges pairs; CharacterRM scores KE, KA, KH, PB, and PU; five PIE criteria are rated with three LLMs and five researchers. Ablations of normalization, number of attributes, quality of G, beta, and model scale are added.

Sample: The main experiments cover two base models and three benchmarks; the PDF does not publish the exact number of valid examples per model, method, and cell. The human evaluation declares 100 samples for the general scenario and 100 for specific personality, rated by five researchers over two days, but does not explain how those ratings are aggregated into the decimal percentages of each comparison. The scale tests add Qwen2.5-3B and 14B; the DSP figure uses 200 examples.

Findings

  • PDD obtains the highest CharacterRM average among open methods: 2.85 on Qwen and 2.81 on LLaMA.
  • The average margin over the best baseline is small: 0.02 for Qwen and 0.06 for LLaMA.
  • Pairwise comparisons usually show more wins than losses, although several win rates do not reach 50%.
  • PDD reaches a Big Five average of 4.57 on both models.
  • Table 3 contradicts per-trait superiority: PDD does not lead four traits with Qwen nor extraversion/neuroticism with LLaMA.
  • Normalization improves the CharacterRM averages from 2.80 to 2.85 and from 2.71 to 2.81.
  • A degraded G reduces the top-5 overlap from 3.97 to 3.66, without intervals or external importance labels.
  • Beta=0.5 yields 44.5% wins, more than beta=1 used in the main experiments.
  • PDD averages 2.85 on Qwen2.5-3B and 3.26 on Qwen2.5-14B, above the reported baselines.
  • The rebuttal reports 2.85 versus 2.76 for dynamic prompting, with 47.0% wins and 33.5% losses.
  • Human and LLM judges rate the PIE rankings as plausible under five criteria created by the authors.
  • One case demonstrates that modulation can reinforce violent content consistent with a profile, not only style or innocuous fidelity.

Limitations

  • A single-sequence log-probability is not an entropy nor proves a CMI estimate.
  • G is generated with the full profile and then used to decide which attributes to reinforce, creating a circular and self-consistent proxy.
  • The theoretical test assumes an unknown monotonic function, bounded iid noise, and a favorable signal-to-noise ratio that is not estimated.
  • The proposed h also depends on the base probability p, although it is handled as a function only of the quotient t.
  • The DSP plot does not validate person importance and reports no coefficient, interval, or fit.
  • The reward in Equation 9 uses a two-token window and invokes y_0 on the first step without a defined convention.
  • The equations and the algorithm omit the dependence of R_norm on the candidate token; as written, the reweighting cancels out.
  • It is not defined how to treat negative importances or rewards, top-k ties, or an L2 norm equal to zero.
  • Cauchy-Schwarz does not guarantee that token-by-token rewards can be proportionally aligned with importance nor that they preserve their ranking.
  • There is no code to resolve the mathematical ambiguities or reproduce PDD.
  • The five PIE metrics are ad hoc, measure visible plausibility, and are not validated against causal contribution or ground truth.
  • Judges see the target profile or trait, favoring explicit and stereotyped lexical expression.
  • PERSONALITYBENCH uses descriptions generated with ChatGPT and keywords of unspecified extraction.
  • Direct scores are close to the ceiling and the mean CharacterRM margins are small.
  • The claim of consistent superiority across five traits contradicts Table 3 itself.
  • The PDF declares p<0.05 without test, unit, pairing, exact p, effect, interval, or correction; the rebuttal limits significance to individual comparisons.
  • Denominators per cell, failures, ties, seeds, independent runs, and model and judge snapshots are missing.
  • The human evaluation does not report independence from the authors, demographics, remuneration, blinding, order, inter-rater agreement, or uncertainty.
  • It is not explained how 100 samples and five evaluators produce the decimal percentages of all comparisons.
  • Top-2 requires approximately three distributions per token; there is no reproducible latency, throughput, memory, energy, or monetary cost.
  • The dynamic prompting, matched compute, CoSER, and safety results appear partially or exclusively in the rebuttal without a frozen artifact.
  • The benchmarks concentrate on fictional/Chinese characters and explicit Big Five descriptions.
  • Populations, elections, diffusion, social networks, or correspondence with real human outcomes are not evaluated.
  • Adherence may amplify violent, manipulative, or stereotyped profiles; there is no safety or bias evaluation.
  • The claim of filtering sensitive attributes cannot be audited and the displayed profiles retain gender, age, and other identities.
  • The official package contains only manuscript sources and figures; processed data, outputs, probabilities, rankings, judgments, and notebooks are missing.
  • Exact versions, templates, runtime, precision, serialized prompts, parsers, retries, and error policy are not published.
  • There are no specific licenses or governance for the subsets, processed profiles, generated descriptions, or human evaluation.

What the study does not establish

  • It does not establish that the log-ratio of a self-generated response is conditional mutual information.
  • It does not establish an external, causal, or human person importance.
  • It does not resolve the circularity of using a profile-conditioned response as a proxy for truth.
  • It does not publish an unambiguous and executable token-by-token policy.
  • It does not demonstrate that PDD outperforms every baseline on every Big Five trait.
  • It does not demonstrate that the small average improvements are robust to judges, seeds, versions, or multiple comparisons.
  • It does not validate the five PIE metrics or the reliability of their judges.
  • It does not demonstrate internal, stable, latent, or human-like personality.
  • It does not demonstrate validity for sociological simulation or prediction of population outcomes.
  • It does not demonstrate that the additional cost is light, efficient, or justified in deployment.
  • It does not demonstrate safety for harmful profiles, minors, sensitive attributes, or stereotypes.
  • It does not allow reproducing results, reconciling the equations with execution, or verifying the rebuttal without code and data.
  • It does not convert experiments described only in OpenReview replies into frozen and reproducible final evidence.

Traceability

Scope: Full text

Version: ICLR 2026 Poster, OpenReview submission 16516; arXiv:2603.01438v1 submitted 2026-03-02; no author-controlled public code/data artifact identified

Consulted source: https://arxiv.org/abs/2603.01438

Review: Codex 30-page visual full-text, TeX/source, OpenReview peer-review, CMI/proxy, token-policy, measurement, statistical, efficiency, safety, artifact and reproducibility audit, 2026-07-18

Approval: Codex fidelity pass, 2026-07-18

English translation: approved, 2026-07-18

Models evaluated

  • Qwen2.5-7B-Instruct, revisión y runtime no especificados
  • LLaMA-3-8B-Instruct, revisión y runtime no especificados
  • Qwen2.5-3B-Instruct, análisis adicional
  • Qwen2.5-14B-Instruct, análisis adicional
  • GPT-4o como juez y comparador cerrado, snapshot no especificado
  • GPT-5 como juez PIE, snapshot no especificado
  • DeepSeek-R1 como juez PIE y comparador cerrado, revisión no especificada

Instruments and metrics

  • Persona Importance Estimation por log-ratio de probabilidad con ablación de atributos
  • Persona-Guided Inference-Time Alignment con recompensa multiobjetivo y regularización KL
  • CharacterRM: Knowledge Exposure, Knowledge Accuracy, Knowledge Hallucination, Persona Behavior y Persona Utterance
  • Comparación por pares mediante GPT-4o
  • Escala directa 1-5 de expresión Big Five
  • Cinco métricas PIE definidas por los autores: relevancia, utilidad, cobertura, independencia y consistencia de ranking
  • Evaluación humana por cinco investigadores
  • Ablaciones de normalización, top-k de atributos, calidad de G, beta y tamaño de modelo
  • Wilcoxon signed-rank mencionado solo en el rebuttal, sin resultados completos

Data used

  • CharacterEval: 1.785 diálogos multivuelta y 77 personajes chinos
  • BeyondDialogue: 280 roles chinos, 31 ingleses y 3.552 escenarios
  • PERSONALITYBENCH: 180.000 preguntas abiertas Big Five
  • DSP: 200 ejemplos seleccionados para un gráfico de probabilidad, no ground truth de importancia de persona
  • 100 muestras seleccionadas para evaluación humana general y 100 para personalidad específica, sin IDs ni asignaciones publicados
  • Paquete TeX oficial con manuscrito y 16 figuras, sin código, datos ni salidas

Evidence and location

  • Method, equations, algorithm, experiments, tables, prompts, human evaluation, ethics, and cases: ICLR 2026/arXiv:2603.01438v1, all 30/30 PDF pages rendered and individually inspected
  • Decision, license, metareview, four reviews, and author responses: Official OpenReview forum lVE8H8QNcx, decision plus all 29/29 replies inspected 2026-07-18
  • Version, date, authorship, ICLR comment, and official package of 23 files: Official arXiv abstract, Atom metadata, and v1 TeX/source archive inspected 2026-07-18
  • Absence of identifiable public repository and implementation: Official paper/OpenReview/source links plus exact-title, arXiv-ID, method-name, author, and authenticated GitHub repository/code searches inspected 2026-07-18
  • Audit of CMI proxy, token policy, metrics, statistics, cost, safety, and reproducibility: reports/verification/article-397-pdd-cmi-proxy-token-reward-claim-statistics-safety-artifact-and-reproducibility-audit.json