Beyond Static Personas: Situational Personality Steering for Large Language Models

Trait induction and control2026arXivApproved editorial review

Authors: Zesheng Wei, Mengxiang Li, Zilei Wang, Yang Deng

Keywords: Personality, Persona conditioning, Activation steering

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

4
Authors
8
Findings
17
Limitations
4
Evidence

Editorial summary

English

Wei and colleagues adapt the psychological idea that a disposition can be expressed differently across situations to LLM steering. Their IRIS method builds, for each Big Five domain and 30 topics, a bank of units whose activation frequency differs by more than ten percentage points under opposite-pole prompts. For a new question, it compares the question's activation pattern with the 30 stored patterns, forms a topic mixture and increases units for the target pole while suppressing opposing ones. On Llama-3-8B-Instruct, one GPT-4o judge scores IRIS 9.59/10 versus NPTI's 9.43 on PersonalityBench and 9.26 versus 9.09 on SPBench; supervised fine-tuning remains one or two hundredths higher. Five psychology graduate students also favor IRIS in the evaluated sample, with 35.4% first-place rankings and a 2.18 mean rank. The evidence supports adaptive control of textual trait expression, not human personality or naturally psychological neurons. SPBench contains 450 synthetic GPT-4o-generated questions; retrieval forces every input into 30 topics, some top-k metrics are near chance and validation relies on another classifier without a documented split. Scores saturate near ten, there are no intervals or repeated runs, most personality poles lower GSM8K or CommonsenseQA point estimates, and no code, dataset, neuron bank, outputs, revisions, seeds or environment are released. The result is therefore promising as a steering technique but is not end-to-end reproducible from the public materials.

Español

Wei y colaboradores adaptan a los LLM la idea psicológica de que una disposición puede expresarse de forma distinta según la situación. Su método IRIS construye, para cada dominio Big Five y 30 temas, un banco de unidades cuya frecuencia de activación difiere más de diez puntos porcentuales entre prompts de polos opuestos. Ante una pregunta nueva, compara su patrón de activación con los 30 patrones históricos, forma una mezcla de temas y aumenta unidades del polo deseado mientras suprime las del contrario. En Llama-3-8B-Instruct, un único juez GPT-4o le asigna 9,59/10 frente a 9,43 de NPTI en PersonalityBench y 9,26 frente a 9,09 en SPBench; el ajuste supervisado queda una o dos centésimas por encima. Cinco estudiantes de psicología también prefieren IRIS en el conjunto evaluado, con 35,4% de primeros puestos y rango medio 2,18. La evidencia respalda control adaptativo de la expresión textual, no personalidad humana ni neuronas psicológicas naturales. SPBench contiene 450 preguntas sintéticas generadas con GPT-4o; la recuperación fuerza toda entrada a 30 temas, algunas métricas top-k quedan cerca del azar y la validación usa otro clasificador sin split documentado. Las puntuaciones saturan cerca de diez, no hay intervalos ni repeticiones, la mayoría de polos reduce GSM8K o CommonsenseQA y no se publican código, datos, banco neuronal, salidas, revisiones, semillas ni entorno. Por ello el resultado es prometedor como técnica de steering, pero no reproducible de extremo a extremo con los materiales públicos.

Research question

Can a neuronal intervention that recovers patterns associated with similar situations control the Big Five expression of an LLM with more precision than prompts and static steering, and do its activations show a consistent thematic dependence?

Method

IRIS uses PersonalityBench questions grouped into 30 themes. For each Big Five domain, theme, and pole, it computes per unit the proportion of output tokens with positive activation; differences greater than ±10 percentage points define positive and negative situational units. At inference it obtains a contrastive vector for the question, computes cosine similarity with the thematic banks, normalizes by softmax, and weights the intervention: it increases positive units according to similarity, gamma, historical 95th percentile, and a sigmoid function, and suppresses negative units above a sigma threshold. It compares with Simple Prompt, P2, ActAdd, NPTI, and LoRA-SFT by means of a GPT-4o judge, human evaluation, ablations, model transfer, general tasks, and latency.

Sample: The bank uses PersonalityBench identification questions distributed across five domains, two poles, and 30 themes; the LDA analysis declares 1,200 questions for each of four selected themes. The automatic evaluation uses approximately 450 questions from the PersonalityBench test and 450 from SPBench, 90 per domain. Five psychology graduate students score the 450 SPBench questions and classify four responses for 300 PersonalityBench sets.

Findings

  • On PersonalityBench, IRIS achieves mean 9.59 and variance 0.30 versus 9.43 and 0.49 for NPTI; it leads the direct baselines on average, but not on Agreeableness or Conscientiousness, and SFT obtains mean 9.61.
  • On SPBench, IRIS obtains 9.26 and 0.58 versus 9.09 and 0.66 for NPTI; SFT remains at 9.27 and Simple Prompt retains the best Agreeableness mean.
  • Human evaluation assigns IRIS 35.4% first places and mean rank 2.18, better than P2, NPTI, and Simple Prompt on the presented set.
  • Automatic results favor IRIS on Qwen3-8B, Gemma-3-12B-IT, and Qwen3 of 0.6B, 8B, and 14B, although gamma changes between families and scores are close to the ceiling.
  • Activation patterns separate by themes in PCA/LDA, evidence of sensitivity to thematic content under contrastive prompts, but not proof of human psychological dynamics.
  • Semantic retrieval is uneven: on PersonalityBench, Agreeableness obtains 6.1% top-2 and 17.2% top-5, practically the random levels of 6.7% and 16.7% for 30 classes.
  • Nine of ten poles fall below the base model on GSM8K and seven of ten on CommonsenseQA; a few poles improve, with no uncertainty that allows distinguishing effect from variation.
  • The source package and the ACL page do not publish implementation, SPBench, neuronal bank, responses, evaluations, or results in processable format.

Limitations

  • The textual expression scored under prompts is not equivalent to human personality, latent trait, identity, or behavioral consistency outside the benchmark.
  • Persona neuron is an operational label based on an arbitrary threshold of ten points; necessity, sufficiency, or causal specificity of each unit is not tested.
  • Thematic differences may capture content, length, tokenization, and generative state, and zeroing out unselected units favors separation in PCA.
  • LDA uses known theme labels without holdout or permutation evaluation; separating thematic texts does not establish a human situation-behavior law.
  • Softmax forces every question into a mixture of 30 themes, with no unknown category, abstention, or out-of-distribution calibration.
  • SPBench reuses the same 30 categories from the bank: it generalizes to new questions, not to new situational categories.
  • The classifier used as semantic reference is trained with all labeled questions and does not document a split; furthermore it is called RoBERTa-Large 325M although the link points to XLM-RoBERTa Large of approximately 0.6B.
  • SPBench is generated and refined by GPT-4o from explicit facets; its high induction validity does not demonstrate ecological or population coverage.
  • The main result depends on a single GPT-4o judge, without human calibration of the scale, alternative judge, repetition, or sensitivity analysis.
  • Scores saturate near the maximum; differences of one or two hundredths lack intervals, tests, effect sizes, or independent replication.
  • Variance is interpreted as stability without precisely defining its unit of aggregation and does not measure stability between runs with greedy decoding.
  • Transfer changes gamma between Llama, Qwen, and Gemma, so it does not keep the entire intervention fixed.
  • The majority of configurations reduces the punctual results of GSM8K or CommonsenseQA; the proposed causal explanations come from cases, not from controlled mediation.
  • Only five evaluators support the human validation; per-person results, data, intervals, or a model treating dependence by evaluator and question are not published.
  • The checklist marks no on consent despite stating in text that it was obtained, leaves ethical review blank, and does not justify wage adequacy relative to the country of the participants.
  • The article attributes ReLU to Llama-3, but the official implementation uses SiLU in a SwiGLU MLP, weakening the comparative description of architectures.
  • Code, dataset, responses, neuronal bank, annotations, judge outputs, model reviews, seeds, dependencies, and reproduction commands are missing.

What the study does not establish

  • It does not demonstrate that LLMs possess human personality, natural psychometric traits, identity, mind, or experience.
  • It does not demonstrate that the selected units are individual personality neurons or that they are free of thematic or linguistic confounding.
  • It does not test a human situation-behavior dynamic; it shows thematic separability and causal change of text within a constructed design.
  • It does not demonstrate open generalization to situations outside the 30 categories nor does it have a mechanism to detect that condition.
  • It does not establish superiority across all domains, metrics, or against SFT; several baselines win specific cells and SFT retains a slightly higher mean.
  • It does not establish that automatic differences near the ceiling are statistically or practically significant.
  • It does not demonstrate that steering preserves reasoning or instruction following; the majority of poles reduces at least one general score.
  • It does not offer a public, licensed, and reproducible SPBench benchmark or an integral run from open artifacts.
  • It does not allow knowing whether all baselines were regenerated under the same environment or whether part of their rows was imported from the prior NPTI work.
  • It does not provide a verifiable reproduction of the results, even though the article has been accepted and published in Findings of ACL 2026.

Traceability

Scope: Full text

Version: arXiv:2604.13846v3; Findings of ACL 2026 final inspected as supplementary authoritative publication

Consulted source: https://arxiv.org/abs/2604.13846

Review: Codex dual 26-page visual full-text, TeX/source, ACL checklist, interface artifact, construct, retrieval, benchmark, statistics, human-study and reproducibility audit, 2026-07-17

Approval: Codex fidelity pass, 2026-07-17

English translation: approved, 2026-07-18

Models evaluated

  • Llama-3-8B-Instruct
  • Qwen3-0.6B
  • Qwen3-8B
  • Qwen3-14B
  • gemma-3-12b-it
  • GPT-4o-2024-08-06 como juez y generador de SPBench
  • Claude-3.5-Sonnet-2024-10-22 como juez de SPBench
  • Clasificador enlazado como FacebookAI/xlm-roberta-large

Instruments and metrics

  • Marco Big Five
  • Probabilidad binaria de activación por unidad FFN
  • PCA y LDA
  • Similitud coseno y recuperación softmax
  • LLM-as-a-Judge de 1 a 5
  • Ranking humano ciego
  • GSM8K
  • CommonsenseQA

Data used

  • PersonalityBench
  • SPBench, 450 preguntas sintéticas no publicadas
  • Taxonomía de 30 temas derivada de UltraChat

Evidence and location

  • Publication, method, results, appendices, prompts, cases, limits, and ethics: Findings of ACL 2026, DOI 10.18653/v1/2026.findings-acl.958, 26 pages rendered and inspected
  • Editable source and figure/interface artifacts, without code or data: arXiv:2604.13846v3 source SHA-256 90637ffa444b5fc8ac299d901dcf175522d8a202c24484563b31846dd416c22d
  • Reproducibility statements, participants, consent, ethics, and AI assistance: Responsible NLP Checklist SHA-256 0a5acb75ccdbf92a0e4d1423818feeaadb6c1f6859b5c40efbfd152aa4e5f97c, 2 pages inspected
  • Construct audit, retrieval, benchmark, statistics, artifacts, and reproducibility: reports/verification/article-363-iris-situational-steering-construct-retrieval-evaluation-artifact-and-reproducibility-audit.json