Synthetic Personality Review

Personality Vector: Modulating Personality of Large Language Models by Model Merging

Trait induction and control2025ACL AnthologyApproved editorial review

Authors: Seungjong Sun, Seo Yeon Baek, Jang Hyun Kim

Keywords: personality modulation, model merging, personality vectors, Big Five traits, continuous control, multidimensional traits, personalized AI

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

3
Authors
35
Findings
92
Limitations
21
Evidence

Editorial summary

English

Sun, Baek, and Kim represent a personality direction as the weight difference between a base model and a variant fine-tuned on one of ten high/low Big Five conditions from Big5-Chat. After training ten variants per backbone, the resulting vectors are scaled or composed with task arithmetic, TIES, and DaRE. For a single trait, GPT-4o-assigned BFI scores from open-ended answers track the scaling coefficient with correlations above 0.9; control weakens when five traits are composed and is partly recovered by DaRE. The full comparison qualifies the claimed advantage: on Llama, simple prompting has the best mean multi-trait BFI correlation (0.834 versus 0.646 for the best merge), while task arithmetic plus DaRE only slightly exceeds prompting on the LIWC composite (0.304 versus 0.289); on Qwen, prompting leads both measures (0.883 and 0.364 versus 0.613 and 0.242). Character, Korean, Chinese, and vision experiments show that weight edits can alter outputs of models sharing compatible Llama components, not general transfer between architectures. Table 3 also contradicts the reversal claim: subtracting the Low vectors leaves all five scores below the base model rather than increasing them. The paper supports controllable scored behavior under its evaluation protocol and presents a useful weight-editing technique; it does not establish stable psychological personality or a universal trait representation.

Español

Sun, Baek y Kim proponen representar una dirección de personalidad como la diferencia de pesos entre un modelo base y una variante ajustada con uno de los diez extremos alto/bajo del Big Five en Big5-Chat. Tras entrenar diez variantes por backbone, esos vectores se escalan o combinan por aritmética de tareas, TIES y DaRE. En el caso de un solo rasgo, la puntuación BFI asignada por GPT-4o a respuestas abiertas sigue el coeficiente con correlaciones superiores a 0,9; el control se degrada al combinar cinco rasgos y mejora parcialmente con DaRE. Sin embargo, la comparación completa matiza la ventaja: en Llama el prompt simple obtiene la mejor correlación BFI media multirrasgo (0,834 frente a 0,646 para la mejor fusión), mientras que task arithmetic + DaRE solo lo supera ligeramente en el compuesto LIWC (0,304 frente a 0,289); en Qwen el prompt domina ambas medidas (0,883 y 0,364 frente a 0,613 y 0,242). Las pruebas con personajes, coreano, chino y visión muestran que los pesos pueden alterar salidas de modelos que comparten componentes Llama compatibles, pero no prueban una transferencia general entre arquitecturas. Además, la Tabla 3 contradice la afirmación de inversión: al restar los vectores Low, las cinco puntuaciones siguen por debajo del modelo base. El trabajo demuestra control de comportamiento puntuado bajo este protocolo y aporta una técnica de edición de pesos interesante; no demuestra una personalidad psicológica estable ni una representación universal de los rasgos.

Research question

Can the difference in weights between a base model and models fine-tuned to extremes of the Big Five function as a scalable and composable vector that continuously modulates the outputs of compatible models, preserves general capabilities, and transfers to personas, other languages, and a vision-language model?

Method

The authors separately fine-tuned Llama-3.1-8B-Instruct and Qwen2.5-7B-Instruct on the ten subsets of Big5-Chat, one per high or low level of each trait, and defined each vector as fine-tuned weights minus base weights. For Llama they scaled the ten vectors with 20 values between 0.1 and 2.0, built 32 multi-trait profiles with four coefficients between 0.1 and 0.4, and tested subtraction with alpha -1. They compared prompting, P² and NPTI with task arithmetic, TIES and DaRE. They measured open-ended responses to the 44 BFI items using a GPT-4o judge, LIWC-22 linguistic composites on 300-word self-presentations, AlpacaEval, MMBench, and cosine similarity of hidden states; they repeated the experiments five times with temperature 0.6. They validated a sample of 400 GPT scores with ten human judges. For transfer they fused vectors with four role agents based on Character-LLM, compatible Llama models in Korean and Chinese, and a VLM with a Llama backbone, and conducted a human preference on responses to PsychoFlicker images.

Sample: Big5-Chat provides 100,000 dialogues, 10,000 for each of the ten high/low conditions of the Big Five. Ten variants per backbone were trained for Llama and Qwen; each fine-tune used two A100 GPUs of 80 GB for approximately 40 minutes. The main single-trait experiment produced 200 Llama fusions and the multi-trait experiment 128 configurations for 32 profiles. The trials were repeated five times with temperature 0.6, although it is not clarified whether that includes new training seeds. The GPT judge validation used 400 responses and ten human annotators (8 men, 2 women, mean age 29.1; 10 USD/hour). The VLM evaluation declares 200 images and ten annotators, but the description of how 20 images from five users per condition were selected is arithmetically ambiguous.

Findings

  • The article was published in the main proceedings of EMNLP 2025, pages 24656-24677, with DOI 10.18653/v1/2025.emnlp-main.1253 and CC BY 4.0 license.
  • The final version of EMNLP and arXiv v1 contain the same textual body; the observed tokenized difference is limited to the proceedings header.
  • The method requires training ten trait models per backbone before fusion can operate without additional training.
  • The fine-tunes used three epochs, sequence 2048, bfloat16, and 40 warm-up steps; Llama used batch 64 and 5e-6, and Qwen batch 32 and 1e-5.
  • Each personality fine-tune ran on two A100 GPUs of 80 GB for about 40 minutes.
  • In Llama, the High models reach BFI 5.00, 4.89, 4.48, 4.69, and 4.38 for O, C, E, A, and N, compared to 4.24, 3.65, 2.80, 4.24, and 2.57 for the base.
  • The Llama Low models reach 2.06, 2.02, 1.95, 1.33, and 2.20 on those five traits.
  • In the single-vector experiment, alpha and the BFI score of the target trait show correlations above 0.9 with p<0.05.
  • LIWC linguistic intensity also tends to increase with alpha, but with lower correlations and more noise than the BFI scored by GPT-4o.
  • Scaling Low Agreeableness up to alpha 2 produces openly hostile and aggressive language, a risk shown by the article itself.
  • The composition of five traits degrades the correlation compared to the single-vector case due to interference between weight differences.
  • In Llama, task arithmetic improves from 0.575 to 0.646 mean BFI correlation when applying DaRE; TIES improves from 0.577 to 0.623.
  • In Llama, the simple prompt retains the best multi-trait BFI correlation, 0.834, above all weight fusions.
  • In the Llama linguistic composite, task arithmetic + DaRE obtains 0.304, barely above the prompt, 0.289.
  • The paper calls the DaRE improvement significant, but presents no statistical test, interval, or error for that comparison in Table 2.
  • The ten weight differences have cosine similarities between 0.30 and 0.55 off the diagonal, compatible with overlap, although they do not by themselves prove the same semantic content.
  • Table 3 contradicts the text of RQ3: when subtracting Low vectors, O, C, E, A, and N remain at 3.44, 3.22, 2.75, 3.60, and 1.75, all below the base and not above.
  • Subtracting High vectors reduces the five scores and generates responses with disclaimers such as 'As an AI', so negation also harms naturalness and conversational capability.
  • In multilingual transfer, 18 of 20 conditions change in the intended direction; Korean High Conscientiousness drops from 4.00 to 3.38 and Chinese Low Neuroticism rises from 1.43 to 2.00.
  • Table 4 omits arrows precisely in those two exceptions, although the prose generalizes that the modulation was successful in the intended direction.
  • The quantitative example of role agents raises Beethoven's Extraversion from 2.4 to 3.9; the rest of the persona evidence is presented mainly through radar charts and examples.
  • The Korean, Chinese, persona, and visual transfer reuses a Llama-compatible backbone or component with the vectors, not an arbitrary transfer between architectures.
  • In the VLM preference, High Openness, High Extraversion, and High Agreeableness receive 100% preference, but High Neuroticism obtains 40% against 60% ties.
  • Low Agreeableness only obtains 55% preference with 45% ties; therefore, Figure 6 does not support a uniform clear preference across the ten conditions.
  • The VLM judges select the response that offers a more favorable impression of the image, not an independent measure that the response correctly expresses the trait.
  • In Qwen, the prompt outperforms the best fusion in both mean BFI (0.883 vs. 0.613) and LIWC (0.364 vs. 0.242).
  • The P² baseline achieves mean BFI correlations of 0.888 in Llama, 0.922 in Qwen, and 0.918 in GPT, showing that a prompt-based description remains highly competitive.
  • The AlpacaEval performance of the single trait is visually stable between 0.1 and 2.0, but the multi-trait composition drops when the sum of coefficients exceeds 2.0.
  • The Korean model shows a small reduction in translated AlpacaEval and the VLM a small reduction in MMBench; the figures provide no numerical tables or uncertainty.
  • The human validation declares a mean correlation between judges of r=0.85 and a GPT-human correlation of r=0.92, both with p<0.05.
  • The hidden state similarity drops in deep layers for BFI, remains above 0.90 when comparing prompting and fine-tuning, and is nearly constant in GSM8K; this is descriptive evidence, not a causal identification of representations.
  • The official repository linked from the PDF has existed since May 2025 and published fusion and interview scripts.
  • The repository version available during publication did not include the training pipeline; this was added on 19 March 2026 in commit c83e619.
  • As of 15 July 2026, the repository contains no LICENSE file, versioned release, models, raw results, or a comprehensive automated run of all the paper's experiments.
  • The Responsible NLP checklist marks recruitment, consent, geography, and use of AI assistants as reported, but the cited section A.3 does not provide those complete details.

Limitations

  • The expression 'without additional training' is only true in the fusion phase.
  • Extracting the vectors requires ten complete fine-tunes per backbone.
  • The study trains twenty trait models between Llama and Qwen before evaluating the fusions.
  • Weight transfer requires compatible parameter shapes and backbones.
  • The Llama vectors do not transfer to Qwen; for Qwen, their own vectors are trained.
  • Transfer between architectures is inferred from similar patterns, not from applying the same vector between Llama and Qwen.
  • The Korean, Chinese, persona, and visual models share the backbone or Llama components used for fusion.
  • Only two open base families of approximately 7-8B parameters are studied.
  • The technique is not tested on larger models or incompatible architectures.
  • The vectors inherit any bias, style artifact, and task confusion from Big5-Chat.
  • It is not separated how much of the vector corresponds to the trait and how much to the style of the fine-tuning dialogues.
  • Vectors are not trained with several independent personality datasets.
  • No replication with multiple training seeds is reported.
  • The claim of five repetitions does not clarify whether the fine-tune or only the stochastic generation is repeated.
  • The main figures do not show error bars despite the five repetitions.
  • No confidence intervals are published for the correlations or scores.
  • No exact p-values are published.
  • No correction for multiple comparisons is reported.
  • A high correlation on deliberately ordered and constructed coefficients does not equate to psychological validity.
  • The BFI scores come from responses generated under the instruction to imagine being a real person.
  • The BFI was validated for human self-report, not for generative models.
  • The open-ended responses do not follow the standard BFI response and scoring procedure.
  • A single GPT-4o judge converts the open-ended responses into scores from 1 to 5.
  • The exact snapshot, date, temperature, or seed of the evaluating GPT-4o is not identified.
  • The evaluation depends on another LLM and may reward obvious lexical markers produced by the fine-tune.
  • The human validation covers 400 responses, but does not explain how they were sampled across traits, models, and coefficients.
  • Ten annotators is a small basis for validating the entire experimental space.
  • The annotator distribution is 8 men and 2 women and only the mean age is reported.
  • No recruitment platform or procedure is reported.
  • No geographic location of the annotators is reported.
  • How consent was obtained is not described.
  • IRB approval is declared without identifying the institution, protocol, or approval number.
  • The mean correlation between judges does not measure absolute agreement and does not replace ICC, error, or calibration.
  • The GPT-human correlation can be high even if there is systematic scoring bias.
  • No confusion matrix, MAE, or distribution of GPT-human disagreements is published.
  • The factorial structure of the Big Five in the generated responses is not tested.
  • Invariance across models, languages, or modalities is not tested.
  • Test-retest reliability on separate dates is not tested.
  • It is not evaluated whether the profile predicts behavior outside of explicit questions and self-presentations.
  • The LIWC composites use five manually selected indicators per trait.
  • No weights or differentiated orientations are published for LIWC indicators that may associate in different directions.
  • It is unclear on which set the LIWC normalization minima and maxima are calculated.
  • A 300-word self-presentation is a narrow behavioral sample.
  • Prompt, P², NPTI, and vectors use different scales and mechanisms, without equalizing the fine-tuning budget or hyperparameter selection.
  • The ChatGPT model and version used to generate the P² descriptions are not precisely identified.
  • The neuron selection and reproducible details of NPTI are referred to PersonalityBench and are not fully documented in this paper.
  • The improvement called significant for DaRE in Table 2 lacks published statistical testing.
  • The simple prompt widely outperforms the fusions in Llama multi-trait BFI.
  • The simple prompt outperforms the fusions in both BFI and Qwen multi-trait LIWC.
  • The LIWC advantage of task arithmetic + DaRE over prompt in Llama is small, 0.304 vs. 0.289.
  • Composition loses tracking capability when overlapping vectors are summed.
  • A multi-trait coefficient sum above 2 degrades AlpacaEval.
  • An independent coefficient is not optimized for each trait of the profile.
  • The interaction between alpha, DaRE drop rate, and TIES trim rate is not systematically analyzed.
  • Table 3 contradicts the interpretation that subtracting a Low vector raises the opposite trait.
  • Subtraction also induces disclaimers and conversational deterioration.
  • The cosine similarity between vectors does not demonstrate that they encode a common personality essence.
  • Orthogonalization or causal isolation of traits is not studied.
  • Two of the twenty multilingual conditions move in the opposite direction to the intended one.
  • Korean AlpacaEval is translated with the GPT API without human validation of equivalence.
  • Psychometric invariance of BFI across English, Korean, and Chinese is not evaluated.
  • Transfer to personas is concentrated on four fictional figures and a few examples.
  • The persona radar charts provide no intervals or aggregate tests.
  • The VLM experiment uses a model with a compatible Llama component and does not test transfer to a VLM from another family.
  • The PsychoFlicker selection phrase, 20 images each from five people per condition, does not match the declared total of 200.
  • The exact list of selected images is not published.
  • The VLM human preference measures favorable impression, not independent correspondence with a psychological trait.
  • High Neuroticism obtains a majority of ties and Low Agreeableness only a narrow preference.
  • Blinding, side randomization, and inter-judge agreement for the VLM task are not described.
  • The traits of Flickr users are used to choose images, but do not validate that each image expresses the assigned trait.
  • AlpacaEval and MMBench are shown as curves without numerical tables, error, or equivalence testing.
  • The visual stability of a curve does not prove the absence of relevant degradation.
  • The hidden state comparison is correlational and descriptive.
  • A cosine similarity above 0.90 between prompting and fine-tuning does not demonstrate equivalent psychological representations.
  • Figure 8c compares states in GSM8K, but does not publish a complete reasoning accuracy evaluation for that intervention.
  • The paper does not perform layer-by-layer causal interventions that localize the supposed trait.
  • Memory, biographical continuity, or longitudinal persistence of a personality is not evaluated.
  • Robustness to paraphrasing, question order, or adversarial prompts is not evaluated.
  • A broad safety evaluation of the modified models is not performed.
  • The study itself observes aggressive language with Low Agreeableness and warns of misuse.
  • The frequency or severity of aggressive behavior is not quantified.
  • The linked repository does not fix a commit in the PDF.
  • The code version existing during publication did not contain the vector training.
  • The training pipeline was added months after EMNLP 2025.
  • The publication code contains local absolute paths, an API key placeholder, and hardcoded CUDA loading.
  • The publication interview script attempts to open config.json, which does not appear in that snapshot.
  • There is no LICENSE file in the official repository.
  • There are no releases, tags, software DOI, or published automated tests.
  • Checkpoints of the twenty trait models or the fusions are not published.
  • Raw responses, human annotations, or complete per-repetition results are not published.
  • The checklist claims that recruitment, consent, geography, and use of AI assistants are reported, but section A.3 does not contain all that information.
  • Generalization to diverse domains relies on three closely compatible extensions and not on independent replications.

What the study does not establish

  • It does not demonstrate that LLMs have an internal human personality.
  • It does not validate the BFI as a psychometric measure of generative models.
  • It does not demonstrate a persistent identity outside the evaluated tasks.
  • It does not demonstrate that each vector contains exclusively one Big Five trait.
  • It does not separate personality from style, prompt obedience, or Big5-Chat distribution.
  • It does not demonstrate transfer of the same vector between Llama and Qwen architectures.
  • It does not demonstrate transfer to models that do not share compatible parameters.
  • It does not demonstrate that fusion outperforms prompting in multi-trait BFI.
  • It does not demonstrate general superiority over P².
  • It does not demonstrate independent or orthogonal composition of the five traits.
  • It does not demonstrate that subtracting a Low vector inverts the trait in the opposite direction.
  • It does not demonstrate that subtraction preserves conversational capability.
  • It does not demonstrate that the twenty multilingual conditions change correctly.
  • It does not demonstrate cultural or linguistic invariance of the scores.
  • It does not demonstrate that VLM preference measures personality rather than favorable tone.
  • It does not demonstrate absence of capability loss through an equivalence test.
  • It does not demonstrate that weight or hidden state similarity identifies a causal mechanism.
  • It does not demonstrate that the control is safe at high scales.
  • It does not offer comprehensive reproducibility of the experiments as published.
  • It does not justify interpreting the results as a universal and generalizable representation of personality.

Traceability

Scope: Full text

Version: EMNLP 2025 main conference paper, ACL Anthology 2025.emnlp-main.1253, pp. 24656-24677, DOI 10.18653/v1/2025.emnlp-main.1253, CC BY 4.0

Consulted source: https://aclanthology.org/2025.emnlp-main.1253/

Review: Codex full-text, peer-reviewed-source, visual, bilingual-fidelity, weight-merging, psychometric-validity, baseline-comparison, transfer-claim, internal-consistency, reproducibility, code-history, safety and checklist-transparency audit, 2026-07-15

Approval: Codex fidelity pass, 2026-07-15

English translation: approved, 2026-07-18

Models evaluated

  • Llama-3.1-8B-Instruct
  • Qwen2.5-7B-Instruct
  • Llama-3.1-Korean-8B-Instruct
  • Llama3.1-8B-Chinese-Chat
  • Llama-3.1-8B-Vision
  • Character-LLM role-playing variants
  • GPT-4o evaluator
  • ChatGPT-generated P² descriptions

Instruments and metrics

  • Big Five Inventory (BFI), 44 open-ended interview items
  • GPT-4o five-point BFI annotation
  • LIWC-22 trait-specific composite features
  • AlpacaEval instruction-following benchmark
  • MMBench vision-language benchmark
  • Human BFI annotation validation
  • Human pairwise image-response preference
  • Pearson correlation
  • Cosine similarity of parameters and hidden states

Data used

  • Big5-Chat
  • Character-LLM
  • PsychoFlicker
  • GSM8K
  • PersonalityBench
  • AlpacaEval
  • MMBench

Evidence and location

  • Bibliographic identity, DOI, pagination, and license: ACL Anthology 2025.emnlp-main.1253; EMNLP 2025; pp. 24656-24677; DOI 10.18653/v1/2025.emnlp-main.1253; CC BY 4.0
  • Complete final PDF inspected: .cache/editorial-sources/article-082/source.pdf; sha256 49c53861a8e76964ba83e60313041d2f70e38761df01e98435d6a5022ec46a10
  • Content equivalence between arXiv v1 and EMNLP: Token-normalized comparison of arXiv v1 and EMNLP PDF: ratio 0.9995948; only inserted proceedings header after boilerplate filtering; audited 15 Jul 2026
  • Definition of vectors, models, and design: Sections 3-4, paper pp. 24658-24660
  • Hyperparameters, computation, and fine-tuning cost: Appendix A.1 and Table 5, paper p. 24669
  • Base, High, and Low results: Table 1, paper p. 24659
  • Single-trait scaling, LIWC, and aggressive language: Figures 2-3 and RQ1, paper pp. 24660-24661
  • Llama multi-trait comparison: Table 2 and RQ2, paper pp. 24661-24662
  • Subtraction contradiction: Table 3 versus RQ3 prose, paper pp. 24662-24663: all Low-subtraction scores are below Base despite claim they increased
  • Korean and Chinese transfer and their two exceptions: Table 4 and Section 5.4.2, paper pp. 24662-24663
  • Visual preferences and VLM examples: Figures 6-7 and Section 5.4.3, paper pp. 24663-24664; Figure 17, p. 24677
  • Hidden state analysis: Figure 8 and Section 5.5, paper p. 24664
  • Human validation and LIWC composites: Appendix A.3 and Table 8, paper p. 24670
  • P² and Qwen baselines: Figures 9-10 and Table 9, paper p. 24672; Figures 14-15 and Table 10, p. 24674
  • General performance and coefficient sensitivity: Figures 11-13, paper p. 24673
  • Role agent results: Figure 5, paper p. 24662; Tables 11-14 and Figure 16, pp. 24675-24676
  • Acknowledged limitations and risks: Limitations and Ethical Considerations, paper pp. 24664-24665
  • Responsibility checklist inspected: .cache/editorial-sources/article-082/supplements/audit/responsible-nlp-checklist.pdf; all 2 pages rendered and checked; sha256 aa3b94fc08ba2e9152840a0b43910b21e4ae3fad25c533f83df4233dcef34025
  • Official repository and state during publication: https://github.com/RSS-researcher/Personality_vector; publication-era commit a60ae5274dd588f907a69c20744a26dddd0d4dac; archived snapshot sha256 abbb9502abf42d9b3d376ea4c17c98c5ac6fd0afb45a8fab8aa830df17a14a01
  • Current code state and late pipeline: Official repository commit c83e619922f9e1ab9bca2d61311f724f4f26ed51 dated 19 Mar 2026; archived snapshot sha256 113d8f6c8c1aad2ab8d45009952b4138af28721a6ddf71660be30247099dd57a
  • Comprehensive reading and visual verification: All 22 proceedings PDF pages and both Responsible NLP checklist pages rendered and inspected, including Tables 1-14 and Figures 1-17; checked 15 Jul 2026