Activation-Space Personality Steering: Hybrid Layer Selection for Stable Trait Control in LLMs

Trait induction and control2026ACL AnthologyApproved editorial review

Authors: Pranav Bhandari, Nicolas Fay, Sanjeevan Selvaganapathy, Amitava Datta, Usman Naseem, Mehwish Nasim

Keywords: Large Language Models, Personality, Big Five, Persona, Personality Control

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

6
Authors
37
Findings
61
Limitations
16
Evidence

Editorial summary

English

This EACL 2026 long paper proposes Activation-Space Personality Steering, an activation intervention designed to shift textual expression along five OCEAN labels without changing model weights. This review uses the definitive ACL Anthology version rather than the 2025 arXiv record and visually inspects all 16 pages. The method produces visible High–Low separations in the published scores, but the evidence does not support several broader abstract claims about a shared psychological subspace, stability, interpretability, or no impact on capabilities.

For each trait, the study takes High/Low Big5-Chat examples, extracts the last non-padding residual state at each layer, computes normalized mean differences, and aggregates layers using non-negative weights whose learning procedure is not specified. The five aggregated vectors are projected through PCA onto three components. One offline layer per trait is selected by combining L2 change, KL, and token-flip diagnostics; a dynamic layer per prompt maximizes logit change. The intervention mixes both layers with heuristic 0.8/0.2 weights, calibrates polarity using labeled prompts, and injects a scaled direction at every decoding step. Intensity α is empirically swept from 4 to 12 and selected so that a purported fluency score remains at least 3.5.

Llama-3-8B-Instruct, Ministral-8B-Instruct-2410, Mistral-Small-24B-Instruct-2501, Qwen2.5-14B-Instruct, and Gemma-3-4B-IT are evaluated at temperature 0.4, top-p 0.95, top-k 50, and repetition penalty 1.1. An unidentified GPT judge rates trait and “fluency” on BFI-style questions and SocialIQA-derived scenarios; judge model, prompt count, repetitions, uncertainty, human agreement, and statistical tests are not reported. High–Low differences derived from Table 1 range from 1.0 to 3.5 points on a 1–5 scale. Llama averages 2.64; in the reported ablation, hybrid exceeds offline-only and dynamic-only, 2.64 versus 1.47 and 0.98. This establishes control over the judge criterion for those prompts, not acquisition of a stable human personality.

The low-dimensionality validation has only five vector-observations per model, one per trait. Three components explain 96.31% for Llama, 96.35% for Ministral-8B, 95.91% for Mistral-24B, and 93.37% for Qwen. The published result is therefore above 90%, not above 95% for every model; Gemma is omitted. There is no comparison with unprojected vectors, other ranks, shuffled labels, or a random basis. That three components summarize five related vectors built from synthetic labels does not establish that personality occupies an intrinsic psychological subspace or that the geometry is causal, stable, or generalizable. The paper also claims multi-trait composition, but experiments score each trait separately.

Quality-retention evidence is particularly weak. The appendix publishes a FLUENCY_TEMPLATE that asks how strongly a response reflects the OCEAN trait and reuses its trait factors; it does not define grammar, coherence, naturalness, or relevance. Without code, it is impossible to know whether the tables used a different prompt, but the published instrument does not validate fluency. Table 1 explicitly computes “variance” across the three High/Base/Low conditions, not across runs, so it cannot measure stability; the narrative nevertheless interprets it as consistency across multiple runs. α is selected using the same evaluation threshold, making fluency preservation partly selection-conditioned.

Knowledge tables contradict “without impact.” Some Llama conditions reduce MMLU by 2.21 points and ARC by 6; Ministral-8B drops by up to 5.42 and 9; Gemma by up to 3.90 and 6. Mistral-24B and Qwen receive no capability evaluation. Ministral Table 4 contains arithmetic inconsistencies: for example, 72.02 to 67.50 is −4.52 points, not −5.05. The 11 MMLU topics are unnamed, aggregation is unspecified, and ARC is limited to 500 questions. Toxicity, bias, misinformation, instruction following, and safety are not tested even though activation manipulation can affect those behaviors.

Overall, the paper offers a plausible architecture and descriptive evidence that directions constructed from Big5-Chat can polarize text from five models under a judge aligned to the same labels. It does not yet establish robust, interpretable, or safe steering: code, artifacts, evaluation sizes, a versioned judge, same-protocol baselines, human validation, statistical tests, repetitions, and out-of-distribution evaluation are missing. The result should be presented as experimental control of High/Low textual expression, not validated personality manipulation or a deployment-ready personalization system.

Español

Este artículo largo de EACL 2026 propone Activation-Space Personality Steering, un método de intervención en activaciones para desplazar la expresión textual de cinco etiquetas OCEAN sin modificar los pesos del modelo. La revisión usa la versión definitiva de ACL Anthology, no el registro arXiv de 2025, e inspecciona visualmente sus 16 páginas. El método produce separaciones High–Low visibles en las puntuaciones publicadas, pero la evidencia no respalda varias afirmaciones más amplias del abstract sobre un subespacio psicológico compartido, estabilidad, interpretabilidad o ausencia de impacto en capacidades.

Para cada rasgo, el estudio toma ejemplos High/Low de Big5-Chat, extrae el último estado residual no-padding por capa, calcula diferencias normalizadas de medias y agrega capas con pesos no negativos cuyo aprendizaje no se especifica. Los cinco vectores agregados se proyectan mediante PCA sobre tres componentes. Una capa offline por rasgo se elige combinando cambio L2, KL y flip del token; una capa dinámica por prompt maximiza el cambio de logits. La intervención mezcla ambas capas con pesos heurísticos 0,8/0,2, calibra la polaridad con prompts etiquetados e inyecta una dirección escalada en cada paso de decodificación. La intensidad α se barre empíricamente entre 4 y 12 y se elige para que una supuesta puntuación de fluidez no baje de 3,5.

Se evalúan Llama-3-8B-Instruct, Ministral-8B-Instruct-2410, Mistral-Small-24B-Instruct-2501, Qwen2.5-14B-Instruct y Gemma-3-4B-IT con temperatura 0,4, top-p 0,95, top-k 50 y repetition penalty 1,1. Un juez GPT no identificado puntúa rasgo y “fluidez” sobre preguntas tipo BFI y escenarios derivados de SocialIQA; no se publican modelo del juez, número de prompts, repeticiones, incertidumbre, acuerdo humano o tests. Las diferencias High–Low derivadas de la Tabla 1 son 1,0–3,5 puntos en la escala 1–5. Para Llama, el promedio es 2,64; en la ablación publicada, hybrid supera a offline-only y dynamic-only, también con 2,64 frente a 1,47 y 0,98. Esto muestra control sobre el criterio del juez en esos prompts, no que el modelo haya adquirido personalidad humana estable.

La validación de baja dimensión usa solo cinco observaciones-vector por modelo: una por rasgo. Tres componentes explican 96,31% en Llama, 96,35% en Ministral-8B, 95,91% en Mistral-24B y 93,37% en Qwen. Por tanto, el dato publicado es “más de 90%”, no “más de 95%” para todos; Gemma se omite de esa tabla. No hay comparación con vectores sin PCA, otros rangos, labels permutados o una base aleatoria. Que tres componentes resuman cinco vectores construidos con etiquetas sintéticas relacionadas no demuestra que la personalidad ocupe un subespacio psicológico intrínseco ni que la geometría sea causal, estable o generalizable. El artículo también dice soportar composición multi-rasgo, pero los experimentos puntúan cada rasgo de forma individual.

La evidencia de preservación de calidad es especialmente débil. El apéndice publica un FLUENCY_TEMPLATE que pide evaluar cuánto refleja la respuesta el rasgo OCEAN y reutiliza sus factores; no define gramática, coherencia, naturalidad o relevancia. Sin código, no puede saberse si las tablas usaron otro prompt, pero el instrumento publicado no valida fluidez. Además, la “varianza” de la Tabla 1 se calcula explícitamente entre las tres condiciones High/Base/Low, no entre ejecuciones, y por eso no mide estabilidad; aun así el texto la interpreta como consistencia entre runs. α se selecciona usando ese mismo umbral de evaluación, de modo que la preservación de fluidez también está condicionada por selección.

Las tablas de conocimiento contradicen “sin impacto”. En Llama, algunas condiciones reducen MMLU 2,21 puntos y ARC 6; en Ministral-8B, las caídas llegan a 5,42 y 9; en Gemma, a 3,90 y 6. No se evalúan capacidades en Mistral-24B o Qwen. La Tabla 4 de Ministral contiene deltas aritméticamente inconsistentes: por ejemplo, 72,02 a 67,50 equivale a −4,52 puntos, no −5,05. Los 11 temas de MMLU no se identifican, no se explica la agregación, y ARC se limita a 500 preguntas. Tampoco se evalúan toxicidad, sesgo, desinformación, seguimiento de instrucciones o seguridad, aunque manipular activaciones puede afectar esos comportamientos.

En suma, el artículo aporta una arquitectura plausible y evidencia descriptiva de que direcciones construidas con Big5-Chat pueden polarizar texto de cinco modelos bajo un juez afín a esas mismas etiquetas. No demuestra todavía steering robusto, interpretable o seguro: faltan código, artefactos, tamaños de evaluación, juez versionado, baselines ejecutados en el mismo protocolo, validación humana, tests, repeticiones y evaluación fuera de distribución. El resultado debe presentarse como control experimental de expresión textual High/Low, no como manipulación validada de personalidad ni como sistema listo para personalización real.

Research question

Can a High–Low activation direction derived from Big5-Chat, compressed with PCA and injected into a combination of an offline layer and a dynamic layer, shift the textual expression of each OCEAN trait across several open LLMs without degrading fluency or general capability?

Method

Final residual states are extracted per layer from synthetic High/Low examples of Big5-Chat and normalized differences are formed per trait. Layers are aggregated with unspecified non-negative weights, five vectors are projected onto three PCA components, and polarity and intensity are calibrated. An offline layer is selected with L2, KL and token flip; another dynamic layer with logit change per prompt; both are injected with a 0.8/0.2 mixture. Five models are evaluated with BFI-type questions and SocialIQA scenarios by an unidentified GPT judge; three models are tested on subsets of MMLU and ARC. Hybrid is compared with offline-only and dynamic-only by means of descriptive scoring differences.

Sample: There are no human participants. The construction units are synthetic Big5-Chat dialogues labeled High/Low; the public dataset has 100,000 rows, but the subset actually used is not determined by the contradictory counts in the method. The evaluation units are questionnaire-type prompts and scenarios derived from SocialIQA whose number is not reported. The capability tests use validation of 11 unidentified MMLU subjects and 500 ARC questions. No number of generations, seeds or repetitions is indicated.

Findings

  • The definitive version is a long paper from EACL 2026, pages 6388–6403, DOI 10.18653/v1/2026.eacl-long.300.
  • The old canonical record had 2025 from the preprint; the definitive publication is from March 2026.
  • The public Big5-Chat contains 100,000 rows: 20,000 per trait and 10,000 per High/Low level.
  • The paper does not allow identifying its subset: it says 20,000 instances and also 5,000 per level for each of five traits.
  • The method constructs directions by High–Low mean difference and does not learn a personality representation from continuous human measures.
  • The offline layer and the dynamic layer are mixed with heuristic weights 0.8/0.2.
  • The appendix reports that a dynamic contribution above 30% reduced steering and capability, but does not publish the curve or all trials.
  • α is selected empirically between 4 and 12 using a score threshold of at least 3.5.
  • The first three PCs explain 96.31%, 96.35%, 95.91% and 93.37% in the four models of Table 2.
  • Qwen contradicts the claim of more than 95% retained variance; the common result is only above 90%.
  • Gemma is evaluated in generation but is omitted from the PCA table.
  • PCA is fitted to a matrix formed by only five aggregated vectors, one per OCEAN label.
  • It is not shown that PCA improves steering over the original directions, nor are other ranks evaluated.
  • The High–Low separations of Table 8 span 1.0–3.5 points on a 1–5 judge scale.
  • Llama obtains separations 1.2/2.8/3.0/3.2/3.0 for O/C/E/A/N.
  • Ministral-8B obtains 2.4/2.0/2.8/3.3/1.0.
  • Mistral-24B obtains 1.8/1.9/3.5/2.6/2.8.
  • Qwen-14B obtains 1.9/2.9/2.6/1.9/2.7.
  • Gemma-3-4B obtains 2.8/2.3/3.5/2.7/2.7.
  • In the Llama ablation, hybrid averages 2.64 versus 1.47 offline-only and 0.98 dynamic-only.
  • In the Gemma ablation, hybrid averages 2.8 versus 2.06 and 1.8.
  • No tests are published that support the use of significantly for those differences.
  • The comparison with prompt, SFT and DPO takes numbers from another work and another scoring method, not controlled runs in this study.
  • The published fluency prompt evaluates trait reflection and does not contain linguistic fluency criteria.
  • The variance of Table 1 is computed between High/Base/Low and not between repetitions.
  • The text improperly interprets that variance between conditions as stability across multiple runs.
  • The Llama conditions lose up to 2.21 MMLU points and 6 ARC.
  • The Ministral-8B conditions lose up to 5.42 MMLU points and 9 ARC.
  • The Gemma conditions lose up to 3.90 MMLU points and 6 ARC.
  • Mistral-24B and Qwen do not receive capability retention evaluation.
  • Table 4 contains deltas inconsistent with its accuracies; 72.02 to 67.50 is −4.52, not −5.05.
  • The GPT model used as judge is not released, so the scores are not reproducible or comparable over time.
  • The examples show clear stylistic polarization, including caricaturization of the Low pole of Conscientiousness.
  • The High direction of Neuroticism is called positive steering although semantically it does not mean a desirable outcome.
  • The geometry reveals dependencies between directions; the appendix itself says that orthogonalizing them substantially reduces the effect.
  • The work evaluates each trait separately and does not demonstrate simultaneous composition of the five traits.
  • No code repository or artifacts linked from the definitive publication or through exact searches were found.

Limitations

  • Big5-Chat is synthetic and its High/Low levels come from textual conditioning, not from participants with measured psychometric scores.
  • The direction may capture phrases, stereotypes and templates of the Big5-Chat generator rather than a psychological trait.
  • Contamination between Big5-Chat scenarios and evaluation prompts derived from SocialIQA/SODA is not evaluated.
  • The counts of the activation subset are incompatible and there is no index, split or sampling seed.
  • No separation of a training set for directions, polarity calibration, α selection and final evaluation is done in a verifiable way.
  • The learning of non-negative weights per layer is claimed but not defined with objective, algorithm or hyperparameters.
  • The values of the lambda weights that combine L2, KL and flip in the offline selection are not published.
  • The number and content of neutral prompts used in the offline selection are not reported.
  • The size of Pcal and the human procedure for semantic verification of polarity are not reported.
  • The paper says human validation for α without describing participants, judges, protocol, agreement or ethical approval.
  • The dynamic layer is chosen by maximum logit change, a measure of sensitivity that does not guarantee correct semantic direction.
  • Maximizing change may select layers that alter the output a lot for reasons unrelated to the trait.
  • The 0.8/0.2 mixture is heuristic and is selected with the same results that are later presented as ablation.
  • The claim of reproducibility by absolute scaling ignores norm, architecture and prompt differences that the text itself acknowledges.
  • The cost of running per-layer diagnostics and forward hooks at each decoding step is not reported.
  • The PCA rank is fixed at three without sensitivity analysis or out-of-sample validation.
  • With five input vectors, any rank analysis is severely limited and does not estimate a population of traits.
  • PCA conventions are not centered or documented in detail, despite interpreting percentages as psychological structure.
  • Qwen falls below the >95% threshold claimed in contributions and conclusion.
  • Gemma is missing from the PCA analysis and Table 9 does not identify which model the shown geometry comes from.
  • Similarity between vectors does not validate correspondence with human Big Five relationships.
  • There are no permuted labels, alternative dataset, style control or random directions as negative controls.
  • No-PCA is not compared with PCA nor is it quantified what signal is lost when projecting.
  • Selective orthogonalization is described in the appendix but is not formalized nor clearly integrated into the equations of the method.
  • Simultaneous multi-trait steering is not evaluated despite being claimed as an advantage.
  • The number of BFI questions and SocialIQA scenarios is not reported.
  • No complete individual prompts are published except for a few examples.
  • It is not reported how many generations are produced per prompt, condition, trait and model.
  • No seeds are fixed nor is sampling variability with temperature 0.4 reported.
  • The GPT judge has no model name, snapshot, provider, parameters, system prompt or repetition.
  • Positional bias, self-consistency or judge sensitivity is not evaluated.
  • There is no independent human evaluation of trait scores or fluency.
  • The published FLUENCY_TEMPLATE does not ask about fluency; it asks about personality again.
  • The fluency criterion used to calibrate α is therefore invalid or, if the code differed, it is not documented.
  • The evaluation and calibration share judge and threshold, favoring results that pass the selected criterion.
  • The variance between High/Base/Low measures dispersion of the effect and not temporal or between-run stability.
  • There are no standard deviations per repetition, intervals, defined bars, bootstrap or tests.
  • The term significantly is used without hypothesis, statistic, p-value, interval or multiple correction.
  • Equivalence or non-inferiority of fluency/capability versus baseline is not tested.
  • External comparisons with prompt/SFT/DPO use different scoring protocols.
  • The High–Low direction may increase extremes and caricatures without preserving naturalness or utility.
  • Negative examples include stereotyped disorganized, hostile or anxious behaviors, not a nuanced measurement of the trait.
  • Factual content, relevance, length, lexical diversity, repetition or human preference are not evaluated.
  • MMLU is described citing MMLU-Pro but the linked resource and the name used are MMLU; the exact variant is ambiguous.
  • The 11 MMLU subjects and the aggregation method are not listed.
  • ARC uses only 500 questions and no selection, seed or whether all available items are used is reported.
  • Accuracies have no intervals and some differences may be sampling noise.
  • Table 4 contains arithmetic errors in the deltas versus baseline.
  • Drops of up to 9 ARC points are material although the text calls them minor variation.
  • Only three of five models receive capability tests.
  • There are no benchmarks for instruction following, safety, toxicity, bias, truthfulness or privacy.
  • Generalization to other languages, domains, cultures, lengths or dialogue types is not tested.
  • Stability against paraphrases, adversaries, jailbreaks, system prompts or multi-turn conversations is not tested.
  • No code, vectors, configurations, outputs or per-example results are published.
  • Without artifacts it cannot be verified whether the fluency prompt in the appendix matches the actual execution.
  • There is no evaluation of memory/latency cost or compatibility with optimized serving.
  • The interventions require access to activations and are not applicable to closed APIs.
  • The validity of the Big Five in humans does not automatically transfer psychometric validity to scores of text generated by LLMs.
  • The impact of steering on users, persuasion, trust or anthropomorphism is not studied.
  • The filters, moderation or policies recommended in the ethics section are not implemented.
  • Defense funding is declared, but specific dual-use cases are not discussed beyond general warnings.

What the study does not establish

  • It does not establish that an LLM possesses or acquires human personality.
  • It does not demonstrate that the directions represent psychometrically valid Big Five constructs.
  • It does not demonstrate that personality intrinsically occupies a three-dimensional subspace.
  • It does not demonstrate more than 95% retained variance in all models.
  • It does not demonstrate that PCA is necessary or better than using unprojected directions.
  • It does not demonstrate robust simultaneous composition of multiple traits.
  • It does not demonstrate causal interpretability of layers, vectors or logit changes.
  • It does not demonstrate stability between runs with the published variance.
  • It does not demonstrate preservation of fluency with the published prompt.
  • It does not demonstrate absence of general capability degradation.
  • It does not demonstrate superiority over prompting, SFT or DPO under a common protocol.
  • It does not demonstrate generalization to models outside the five architectures, especially closed APIs.
  • It does not demonstrate safety, fairness, absence of toxicity or resistance to malicious use.
  • It does not demonstrate robustness to domains, cultures, languages, prompts or real conversations.
  • It does not offer a reproducible or production-ready system without code, artifacts and validated evaluation.

Traceability

Scope: Full text

Version: EACL 2026 long paper, ACL Anthology 2026.eacl-long.300, DOI 10.18653/v1/2026.eacl-long.300, pages 6388–6403; 16 pages; arXiv:2511.03738v2 retained as superseded source

Consulted source: https://aclanthology.org/2026.eacl-long.300.pdf

Review: Codex definitive-version reconciliation, complete bilingual full-text fidelity pass, all-page PDF visual inspection, Hugging Face Dataset Viewer verification, table arithmetic audit, evaluator-prompt audit, construct-validity and capability-retention review; summaries written from the full paper and source data rather than abstract keywords, 2026-07-15

Approval: Codex fidelity pass, 2026-07-15

English translation: approved, 2026-07-18

Models evaluated

  • Llama-3-8B-Instruct as primary steering model; exact repository identifier not stated
  • mistralai/Ministral-8B-Instruct-2410
  • mistralai/Mistral-Small-24B-Instruct-2501, labeled Ministral-24B in tables
  • Qwen/Qwen2.5-14B-Instruct
  • google/gemma-3-4b-it
  • Unidentified GPT-based evaluation model used as trait and purported fluency judge

Instruments and metrics

  • Last non-padding residual-state extraction by decoder layer
  • Normalized High-minus-Low mean activation directions
  • Unspecified non-negative trait-specific layer weights summing to one
  • Rank-three PCA projection over five aggregated trait vectors
  • Offline layer score combining L2 probability shift, KL divergence and argmax token flip rate
  • Dynamic layer selection by steered-versus-base logit L2 norm
  • Heuristic 0.8 offline and 0.2 dynamic layer mixture
  • Polarity calibration by KL followed by semantic prompt checks
  • Global steering gain 8.0 and per-trait alpha values from 4 to 12
  • GPT-based 1–5 trait judge prompt
  • Published FLUENCY_TEMPLATE that actually asks for trait reflection rather than fluency
  • MMLU over validation sets of 11 unnamed topics
  • ARC-Challenge over 500 questions
  • Descriptive High–Low trait separation and baseline-referenced score changes without inferential tests

Data used

  • wenkai-li/big5_chat: official Hugging Face train split has 100,000 synthetic dialogues, 20,000 per trait and 10,000 per High/Low level
  • Unspecified subset of Big5-Chat; paper simultaneously reports 20,000 instances and 5,000 per High/Low level for each trait
  • Unspecified set of Big Five Inventory-style interview questions
  • Unspecified number of situational questions constructed from SocialIQA
  • MMLU validation subsets for 11 unnamed topics
  • ARC-Challenge subset of 500 questions
  • No released code, activation vectors, prompt list, generated outputs, judge outputs, seeds or result tables located

Evidence and location

  • Definitive publication, year, venue, DOI and pages: ACL Anthology record 2026.eacl-long.300 and definitive PDF page 1
  • Extraction, aggregation and PCA of directions: Definitive paper Sections 3.2–3.3, page 4
  • Offline, dynamic selection and 0.8/0.2 mixture: Definitive paper Section 3.4, page 5
  • Polarity, gain and intensity: Definitive paper Sections 3.5–3.6 and 5, pages 6–7
  • Real Big5-Chat: 100k, 20k per trait and 10k per level: Hugging Face Dataset Viewer API /size and /statistics for wenkai-li/big5_chat on 15 July 2026
  • Contradictory counts of the Big5-Chat subset: Definitive paper Section 3.2, page 4
  • Models and decoding: Definitive paper Results, page 7
  • Exact PCA variance and Qwen 93.37%: Definitive paper Table 2, page 7
  • Trait scores, fluency scores and variance between conditions: Definitive paper Table 1 and Section 5.2, pages 7–8
  • Hybrid/offline/dynamic ablations: Definitive paper Figure 4 and Appendix Tables 6–7, pages 9 and 14
  • High–Low differences and supposed fluency shifts: Definitive paper Appendix Table 8, page 15
  • MMLU and ARC drops and arithmetic deltas: Definitive paper Tables 3–5, pages 8 and 14; independent subtraction from displayed base accuracies
  • Fluency prompt that re-evaluates the trait: Definitive paper Appendix I FLUENCY_TEMPLATE, page 16
  • Declared limitations and ethics: Definitive paper Limitations and Ethical Considerations, pages 9–10
  • Absence of linked code or artifacts: Definitive paper, ACL Anthology record and exact-title/code repository searches checked 15 July 2026
  • Visual inspection: All 16 pages of the definitive EACL 2026 PDF rendered and visually inspected on 15 July 2026