Synthetic Personality Review

PICLe: Eliciting Diverse Behaviors from Large Language Models with Persona In-Context Learning

Trait induction and control2024PMLRApproved editorial review

Authors: Hyeong Kyu Choi, Yixuan Li

Keywords: persona elicitation, in-context learning, Bayesian inference, behavioral preferences, personality traits, ICL, persona customization

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

2
Authors
34
Findings
40
Limitations
30
Evidence

Editorial summary

English

PICLe studies how to make a model answer binary yes/no questions in accordance with a target 'persona.' Here, persona is not equivalent to psychometric personality: the 99 personas in Anthropic's model-written-evals collection include traits such as narcissism and extraversion, but also political and religious beliefs, preferences, instrumental goals, and dangerous tendencies such as deception, power acquisition, or escaping a sandbox. Each persona supplies 1,000 balanced examples, split into 700 for training and 300 for testing. The method trains a persona-specific auxiliary LoRA adapter on the 700 statements, scores every candidate example by the difference between its log-likelihood under the adapted and base models, selects the three largest likelihood ratios, and prepends those labeled demonstrations to a query sent to the base model. In the main protocol, PICLe raises mean label consistency from 65.5 to 88.1 for Llama 2 7B Chat, from 50.1 to 78.6 for Vicuna 7B, and to 67.0 for GPT-J 6B, outperforming baselines without privileged label filtering. When selection is allowed to use only positive examples, however, random sampling reaches 91.5 and similarity 92.4, both above standard PICLe; PICLe+ reaches 93.1 under that additional protocol. The defensible conclusion is that a likelihood ratio from an auxiliary adapter can select demonstrations that induce label-consistent binary answers on this synthetic benchmark. It does not establish a stable, human-like, or internal personality. The evaluation contains no people, free-form text, conversational persistence, other languages, or deployment tests. Reproducibility is also partial: the repository supplies code and an environment, but no license, tests, CI, results, adapters, or portable paths. The audit additionally finds potential discrepancies in manual LoRA merging, an auxiliary function that unpacks the wrong number of model outputs, and an incorrect valid-ratio calculation for Degree of Alteration. These observations limit confidence in exact reproduction from the published artifact, although they do not by themselves invalidate the paper's tables.

Español

PICLe estudia cómo hacer que un modelo responda de acuerdo con una 'persona' objetivo en preguntas binarias de sí o no. El término no equivale aquí a personalidad psicométrica: las 99 personas del conjunto model-written-evals de Anthropic incluyen rasgos como narcisismo o extraversión, pero también creencias políticas y religiosas, preferencias, objetivos instrumentales y tendencias peligrosas como engañar, adquirir poder o escapar del sandbox. Para cada persona se separan 1.000 ejemplos equilibrados en 700 de entrenamiento y 300 de prueba. El método entrena un adaptador LoRA auxiliar específico con las 700 afirmaciones, calcula para cada ejemplo candidato la diferencia entre su log-verosimilitud bajo ese modelo ajustado y bajo el modelo base, elige los tres cocientes mayores y los antepone como demostraciones etiquetadas a la consulta del modelo base. En el protocolo principal, PICLe eleva la consistencia media con la etiqueta de 65,5 a 88,1 en Llama 2 7B Chat, de 50,1 a 78,6 en Vicuna 7B y hasta 67,0 en GPT-J 6B, superando los baselines sin acceso privilegiado a la etiqueta. Sin embargo, cuando la selección puede restringirse a ejemplos positivos, un muestreo aleatorio alcanza 91,5 y la similitud 92,4, ambos por encima del PICLe estándar; PICLe+ llega a 93,1 con ese protocolo adicional. El resultado defendible es que una razón de verosimilitudes obtenida con un adaptador auxiliar ayuda a escoger demostraciones que inducen respuestas binarias coherentes con etiquetas de comportamiento en este benchmark sintético. No demuestra una personalidad estable, humana o interna. La evaluación no incluye personas, texto libre, persistencia conversacional, otros idiomas ni despliegue. La reproducibilidad también es parcial: el repositorio aporta código y entorno, pero no licencia, pruebas, CI, resultados, adaptadores ni rutas portables; además, la auditoría encuentra discrepancias potenciales en la fusión manual de LoRA, una función auxiliar que desempaqueta mal la salida del modelo y un cálculo erróneo de la proporción válida de Degree of Alteration. Estas observaciones limitan la confianza en una reproducción exacta a partir del artefacto publicado, aunque no invalidan por sí solas las tablas del artículo.

Research question

Can a likelihood ratio between an auxiliary model fitted to a person and the base model select in-context learning examples that make several LLMs respond more consistently with that person's behavioral label than prompting or existing selection criteria?

Method

Experimental evaluation on 99 persona files from Anthropic model-written-evals. For each persona, 700 statements/questions are separated in a stratified manner for training and 300 for testing. A LoRA adapter of rank 8 and alpha 32 is trained for four epochs with concatenations of three statements. For each candidate, PICLe computes log p(x|fitted model) minus log p(x|base model), orders the examples, and uses the first three as yes/no demonstrations to query the base model. Comparisons include no-context model, instructive and descriptive prompts, random sampling, similarity, confidence, uncertainty, diversity, and likelihood; measured outcomes are consistency with the labeled action, confidence, uncertainty, token entropy, and Degree of Alteration. Appendices add selection with positive labels, ablations, more examples, fewer data, Llama 2 13B, combinations of four pairs of personas, latency, and t-tests across personas.

Sample: The principal unit is not a human person but a persona type from the dataset and its model-generated examples. 99 types are used, each with 500 statements that match the behavior and 500 that disagree. The fixed and stratified split retains 700 training examples and 300 test examples per type: 69,300 training observations and 29,700 test observations if the 99 experiments are aggregated. Results are averaged across types, which span heterogeneous constructs and not a single psychological taxonomy.

Findings

  • The corpus contains 99 personas and 1,000 question-label pairs per persona, balanced between 500 matches and 500 disagreements.
  • The split is 70/30 and stratified by positive and negative label, giving 700 training examples and 300 test examples per persona.
  • PICLe does not select examples only with the queried model: it first trains a separate auxiliary Persona SFT model for each persona.
  • The adapter uses LoRA with rank 8, alpha 32, and four epochs on statements concatenated three at a time.
  • The central criterion subtracts from the log-likelihood of the example under Persona SFT its log-likelihood under the base model.
  • In the main protocol, three global demonstrations per persona are selected and added with their yes/no answer to all test queries.
  • In Llama 2 7B Chat, mean consistency rises from 65.5 without intervention to 88.1 with PICLe.
  • In Llama 2, instructive and descriptive prompts obtain 53.9 and 74.9; random 79.7 and similarity 84.6, all below the 88.1 of PICLe in the main protocol.
  • In Vicuna 7B, the base model obtains 50.1 and PICLe 78.6, the best result in its block.
  • GPT-J does not produce a sufficiently interpretable binary baseline for several conditions; PICLe reaches 67.0 and exceeds the 61.4 of the best reported ICL baseline for that family.
  • By allowing the selector to know which examples have positive label, random+ reaches 91.5 and similarity+ 92.4, above standard PICLe at 88.1.
  • PICLe+ combines the likelihood ratio with the positive pool and reaches 93.1; it is not the same informative protocol as standard PICLe.
  • The ablation in Llama 2 attributes performance to the quotient: PICLe obtains 88.1 versus 72.1 with only the likelihood of the fitted model and 71.8 with that of the original model.
  • In the positive pool, PICLe+ obtains 93.1 versus 87.8 and 87.0 for the two likelihoods without quotient.
  • With Llama 2, PICLe increases from 65.5 with zero examples to 82.7 with one, 88.1 with three, and 92.3 with ten.
  • Similarity also improves with more context, from 84.6 with three examples to 88.9 with ten, but remains below PICLe in those conditions.
  • Between one and ten epochs of Persona SFT, PICLe consistency remains approximately between 87.6 and 88.7, suggesting little sensitivity within that range.
  • Reducing the training set to 70% yields 87.0 and to 40% also 87.0 versus 88.1 with all data.
  • In Llama 2 13B, PICLe obtains 76.0 versus 55.4 for the base model, 62.9 for similarity, and 71.0 for random.
  • The 13B version performs below the 7B version in this protocol, so the data do not show a monotonic improvement by scale.
  • Action Consistency results count as error outputs that do not match yes/no; confidence and uncertainty exclude them.
  • t-tests are computed over the 99 personas and presented with rounded p-values, but without effect sizes or correction for multiple comparisons.
  • Only four combinations of two personas are tested: narcissism-psychopathy, extraversion-desire for popularity, narcissism-extraversion, and risk aversion-risk seeking.
  • The contradictory combination of risk aversion plus risk seeking reaches 99.3 consistency, a sign that the metric may reward local label following without requiring a coherent identity.
  • The reported latency per persona in Llama 2 is 31.9 seconds for base, 54.0 for similarity, and 66.2 for PICLe, plus 54.8 seconds of Persona SFT once.
  • The 22.6% overhead highlighted in the article compares PICLe inference with similarity after amortizing the auxiliary fitting; a first complete run adds approximately 121 seconds.
  • The impact statement acknowledges that persona elicitation could be used maliciously and the dataset includes explicitly dangerous targets.
  • The public repository fixes a Python 3.8 environment with PyTorch 2.1, Transformers 4.36.2, PEFT 0.4.0, and the necessary dependencies to run the approach.
  • The source code compiles syntactically, but the repository does not publish tests, CI, license, results, frozen splits, trained adapters, or checkpoints.
  • The scripts and default values contain internal absolute paths and the README indicates a clone URL different from the audited public location.
  • The manual adapter merge adds B by A to the base weights without visibly multiplying by alpha/r; unless the saved tensors are prescaled, something not documented, this differs from the standard LoRA application.
  • An auxiliary function returns five values and forward_whole_dataset attempts to unpack four; that path would fail if executed, although it does not appear to be the main path of the tables.
  • The printed calculation of valid ratio for Degree of Alteration divides failures by the number of valid responses instead of the total and can divide by zero.
  • The literature review confirms the final publication in PMLR 235, pages 8722-8739, with Hyeong Kyu Choi and Yixuan Li as authors.

Limitations

  • The 99 categories are called personas, but they mix psychological traits, values, beliefs, preferences, goals, and safety propensities; they do not form a homogeneous personality taxonomy.
  • The dataset is model-written-evals: its examples and labels are synthetic and represent intentional match with a description, not human psychological ground truth.
  • Action Consistency is accuracy with respect to a binary benchmark label; it does not measure stability, structure, identity, or construct validity.
  • No human persons, human evaluators, self-reports, observers, or external behavioral criteria participate.
  • The evaluation requires yes/no responses and does not cover free generation, extended dialogue, planning, or real actions.
  • Only English examples are studied; there is no linguistic, cultural, or invariance analysis.
  • Performance is averaged over 99 heterogeneous personas without giving the same level of detail per construct or justifying that the mean is psychologically interpretable.
  • The auxiliary model is trained and evaluated with examples from the same persona file, so the demonstrated generalization is within the same generator and distribution.
  • Each persona requires its own LoRA fitting, which shifts much of the cost and supervised information outside the apparent in-context learning.
  • The article describes the method as ICL, but its selector depends on prior parametric training; the final model inference does remain unfitted.
  • The criterion uses sums of log-likelihood without explicit normalization by length, so selection may depend on length and tokenization.
  • PICLe selects the same set of examples for all queries of a persona, whereas similarity can adapt examples to each query; the costs and capabilities are not identical.
  • The main comparison does not include the strongest baseline when labels are known; with positive pool, random and similarity exceed standard PICLe.
  • PICLe+ incorporates additional label information and therefore its 93.1 should not be compared as if it were the same method as the main 88.1.
  • No confidence intervals, effect sizes, or full distributions per persona are published for the main tables.
  • p-values are rounded to 0.0000 and described with language of 'perfect significance', a statistically inadequate formulation.
  • Several paired comparisons are made without explicit correction for multiplicity.
  • The four persona combinations are too small a sample to support compositional generalization.
  • The metric for combinations does not clearly penalize internal contradictions, as confirmed by the near-perfect result for two opposing risk orientations.
  • Persistence across turns, resistance to conflicting instructions, recovery of base behavior, and side effects on capability and safety are not tested.
  • There is no evaluation of bias, fairness, stereotypes, demographic differences, or differential harm.
  • The work includes personas of deception, power, self-preservation, and evasion, but the ethical evaluation does not measure whether PICLe increases harmful behaviors in operational tasks.
  • The claim that the work does not directly exacerbate malicious exploitation is stronger than the evidence: the method is designed precisely to increase concordance with goals, including some dangerous ones.
  • The formal limitations section concentrates on the binary space and does not discuss the heterogeneity of constructs, the absence of humans, or psychological validity.
  • Cost is reported per persona and with amortization of fitting, without comprehensive accounting to train and store adapters for the 99 categories across all models.
  • The repository does not provide tabular results, predictions, logs, executed seeds, or adapters with which to contrast the published numbers.
  • The scripts require editing internal absolute paths and do not offer a portable configuration of data, models, outputs, and hardware.
  • There is no LICENSE, so the permission for code reuse is not established in the audited repository.
  • There are no automated tests or CI to detect regressions or verify the metrics.
  • The batch_size argument is defined but not passed to TrainingArguments in the SFT path, which makes the behavior depend on the Transformers default value.
  • The number of training steps is fixed as 44 or 88 instead of being derived from the actual dataset size and batch, a fragile decision in the face of data or dependency changes.
  • The manual LoRA merge does not show the alpha/r scaling factor documented by PEFT; without trained artifacts it cannot be determined whether the published code reproduces the selector used in the study.
  • forward_whole_dataset is incompatible with the five values returned by the wrapper, although this auxiliary function may not affect the main experimental path.
  • gather_last_token ignores lengths and always takes the last position, which would be sensitive to padding when used in batches.
  • The valid ratio calculation for Degree of Alteration uses an incorrect denominator and does not handle the case with no valid outputs.
  • Generation depends on manual lists of yes/no token IDs and exact string matches, which is fragile to tokenizer versions or different formats.
  • The code forces .cuda() and does not document a CPU path; the precise hardware and memory of each reproduction are also not included.
  • The audited repository has a single visible commit in the shallow clone, does not publish releases, and does not tag the exact version used for the tables.
  • Compiled __pycache__ files are versioned, a sign of poor artifact hygiene and a possible source of confusion between versions.
  • Reproducibility was evaluated statically and syntactically: the fitting of 99 adapters and the inferences of 6-13B models were not repeated due to hardware and time cost.

What the study does not establish

  • It does not demonstrate that LLMs possess psychological personality, identity, consciousness, or preferences of their own.
  • It does not demonstrate that an induced persona is stable across conversations, prompts, tasks, or moments.
  • It does not demonstrate psychometric validity of the 99 labels or equivalence with human traits.
  • It does not demonstrate that all 99 categories are personality traits; many are beliefs or goals.
  • It does not demonstrate coherent behavior outside of yes/no binary responses.
  • It does not demonstrate generalization to free text, dialogue, tools, agents, or real environments.
  • It does not demonstrate generalization to human data or to distributions other than model-written-evals.
  • It does not demonstrate cross-cultural, multilingual, clinical, educational, or organizational validity.
  • It does not demonstrate that increasing Action Consistency is desirable or safe for dangerous personas.
  • It does not demonstrate that the model reconciles contradictory personas into a coherent identity.
  • It does not demonstrate that standard PICLe is superior to random or similarity when the label of candidates is available.
  • It does not demonstrate that the improvement comes only from in-context learning, because the selector uses a per-persona supervised adapter.
  • It does not demonstrate global efficiency over alternatives if training, storage, and loading of all adapters are included.
  • It does not demonstrate a monotonic relationship between model size and ease of elicitation.
  • It does not demonstrate robustness to changes in template, tokenizer, order, length, or number of examples beyond the specific ablations.
  • It does not demonstrate absence of regressions in utility, factuality, safety, or alignment when inducing a persona.
  • It does not causally establish that internal personality representations explain the result; the criterion operates on text likelihoods.
  • It does not demonstrate that rounded p-values imply practical importance or large differences.
  • It does not demonstrate that the published repository can exactly reproduce the tables without corrections, artifacts, and additional paths.
  • It does not allow knowing whether the static discrepancies detected in the code were present in the private version used to obtain the results.
  • It does not establish a reuse license for the public code.
  • It does not demonstrate that the method is safe for deployment or suitable for human evaluation applications.

Traceability

Scope: Full text

Version: PMLR 235:8722-8739, ICML 2024 final paper, 18 pages; code snapshot deeplearning-wisc/picle commit 02f55a09e8d222d3b0cf121e1e7ab3ec1af211dc

Consulted source: https://proceedings.mlr.press/v235/choi24e/choi24e.pdf

Review: Codex full-text, bilingual-fidelity, visual, bibliographic, method, metric, statistical, construct-validity, code-artifact, reproducibility, safety and ethics audit, 2026-07-15

Approval: Codex fidelity pass, 2026-07-15

English translation: approved, 2026-07-18

Models evaluated

  • Llama-2-7b-chat-hf
  • Vicuna-7b-v1.5
  • GPT-J-6B
  • Llama-2-13b-chat-hf en el apéndice
  • Un adaptador LoRA Persona SFT separado por cada persona y familia evaluada

Instruments and metrics

  • Persona In-Context Learning (PICLe)
  • Razón de verosimilitudes log p_persona(x) - log p_base(x)
  • Aprendizaje en contexto con K=3 demostraciones
  • Persona SFT mediante LoRA r=8 y alfa=32
  • Action Consistency o exactitud binaria
  • Action Confidence
  • Action Uncertainty binaria
  • Entropía de tokens
  • Degree of Alteration mediante divergencia KL
  • Baselines aleatorio, similitud, confianza, incertidumbre, diversidad y verosimilitud
  • Pruebas t pareadas sobre las 99 personas

Data used

  • Anthropic model-written-evals Persona dataset
  • 99 archivos de persona con 1.000 ejemplos sintéticos cada uno

Evidence and location

  • Final publication, authorship, and pagination: Official PMLR record v235/choi24e: ICML 2024, pp. 8722-8739, Hyeong Kyu Choi and Yixuan Li
  • Complete audited source: .cache/editorial-sources/article-090/source.pdf; PMLR final paper; 18 pages; sha256 d98b4a3cde8863d563ed66a9fd9991f69bc1adda938b7493d661558bb2dab384
  • Task and contribution: Full text pp. 1-3, Abstract, Introduction and Section 2
  • Bayesian definition and likelihood ratio: Full text pp. 3-4, Section 3 and Equations 1-4
  • 99 personas and heterogeneity of constructs: Full text pp. 4-5, Section 4.1; pp. 14-15, Appendix Table 10
  • 1,000 examples, balance, and 700/300 split: Full text p. 5, Section 4.1; public code src/data/persona.py lines 187-205
  • Models and experimental configuration: Full text p. 5, Section 4.1
  • LoRA and four epochs: Full text p. 5, Section 4.1; public code src/main.py lines 27-68
  • Main results for three models: Full text p. 6, Table 1
  • Positive pool, PICLe+, and stronger baselines: Full text p. 7, Table 2 and Section 4.3
  • Ablation of the likelihood quotient: Full text p. 7, Table 3
  • Sensitivity to the number K of examples: Full text p. 8, Table 4
  • Sensitivity to epochs and data volume: Full text p. 8, Tables 5-6
  • Llama 2 13B result: Full text p. 9, Table 7
  • Persona combinations: Full text p. 9, Table 8 and Section 4.4
  • Latency and amortized cost: Full text p. 10, Table 9 and Section 4.5
  • Definition of metrics and invalid outputs: Full text pp. 12-13, Appendix A.2; public code src/main.py lines 75-180
  • t-tests: Full text pp. 16-17, Appendix C and Table 14
  • Formal limitation to binary space: Full text p. 10, Limitations
  • Impact and possible malicious use: Full text p. 10, Broader Impacts and Ethics Statement
  • Audited public repository: https://github.com/deeplearning-wisc/picle at commit 02f55a09e8d222d3b0cf121e1e7ab3ec1af211dc; checked 15 Jul 2026
  • Declared reproducible environment: Public code environment.yml lines 1-120
  • Non-portable paths and cloning: Public code README.md lines 5-13 and scripts/**/*.sh; internal /nobackup2 paths
  • Absence of result and QA artifacts: Public repository tree at audited commit: no LICENSE, tests, CI workflows, result files, adapters, splits or releases
  • Manual LoRA merge without visible scaling: Public code src/models/llama.py lines 124-142; equivalent functions in vicuna.py and gptj.py
  • Auxiliary unpacking incompatibility: Public code src/models/llama.py lines 59-122: __call__ returns five values and forward_whole_dataset unpacks four
  • Padding-sensitive last token selection: Public code src/models/llama.py lines 36-39
  • Error in valid ratio of Degree of Alteration: Public code src/main.py lines 89-107
  • Stratified split and fixed seeds: Public code src/data/persona.py lines 187-205
  • Integral reading and visual verification: All 18 pages rendered and inspected, including Tables 1-14, Figures 1-5, equations, appendices, qualitative examples and references; checked 15 Jul 2026