The Geometry of Persona: Disentangling Personality from Reasoning in Large Language Models

Evaluation and psychometric validity2025arXivApproved editorial review

Authors: Zhixiang Wang

Keywords: Computer science, Representation (politics), Artificial intelligence, Probabilistic logic, Personality

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

1
Authors
6
Findings
8
Limitations
8
Evidence

Editorial summary

English

This preprint presents Soul Engine, a probing and steering system trained on Qwen2.5-0.5B-Instruct to predict character-level OCEAN labels and alter activations toward a target persona. SoulBench is described as a corpus of sentences grouped by character, but the paper does not identify its sources, characters, training-set size, licenses, or split criteria. Each iteration concatenates three random sentences from one character to encourage stylistic invariance. The alleged psychological ground truth does not come from people or an inventory: Doubao-Seed-1.6 generates the labels from full character profiles. The 24-layer, 896-dimensional base model freezes layers 0–19 but fine-tunes layers 20–23 and normalization components, so the abstract's claim that backbone weights are not modified is incorrect. Two heads consume the final representation: a 256-dimensional identity MLP trained with InfoNCE and a five-output linear probe trained by MSE against teacher labels. A regularizer forces orthogonality among directions in the OCEAN head; orthogonality is therefore imposed by the objective rather than discovered spontaneously. Nor does the loss enforce or measure orthogonality between personality, identity, and reasoning circuits. On 1,000 validation examples, the best MSE is .0113 and final MSE .0118. The paper translates .0113 into “≈99% accuracy,” but MSE and accuracy are not interchangeable, and it reports no MAE, R², correlations, calibration, baselines, intervals, or per-trait or per-character results. Its contribution section also promises MSE < .01, which the table does not reach. A t-SNE of the same 1,000 embeddings, coloured only by teacher-labelled Openness, shows two visual groupings; t-SNE cannot establish orthogonality, global continuity, separation of all five traits, or independence from reasoning. For steering, a normalized vector between target-persona and neutral means is added to the residual stream with strength α. Although the method defines intervention at layer L−1, the ablation scans layers 10–20 and calls layers 14–16 with boost 6–8 optimal. The heatmap metrics “Villainy” and “Sanity” are undefined, with no sample, evaluator, repetitions, or uncertainty, and the caption confuses dark colour with greater adherence even though high Villainy is shown in red. No generation examples, prompting/SFT/LoRA comparisons, new-profile tests, or human evaluations are provided. Most importantly, no reasoning or intelligence benchmark is run, so claims of zero alignment tax, preserved original intelligence, and a subspace distinct from reasoning are untested. This is a preliminary demonstration on one 0.5B model using labels from another LLM. It contributes a possible architecture and an internal regression error, but does not validate psychological personality, deterministic control, capability preservation, or safety. The proposed “Safety Interceptor” and subtraction of malicious vectors remain future work.

Español

Este preprint presenta Soul Engine, un sistema de sondeo y steering entrenado sobre Qwen2.5-0.5B-Instruct para predecir etiquetas OCEAN de personajes y modificar activaciones hacia una persona objetivo. SoulBench se describe como corpus de frases por personaje, pero el artículo no identifica sus fuentes, personajes, número de ejemplos de entrenamiento, licencias ni criterios de partición. En cada iteración concatena tres frases aleatorias del mismo personaje para que el modelo aprenda regularidades de estilo. Las supuestas etiquetas psicológicas no proceden de personas ni de un inventario: Doubao-Seed-1.6 las genera a partir del perfil completo del personaje. El modelo base tiene 24 capas y 896 dimensiones. Se congelan las capas 0–19, pero se ajustan las capas 20–23 y normalizaciones, por lo que es incorrecta la afirmación del abstract de que no se modifican los pesos del backbone. Dos cabezas reciben la representación final: una MLP de identidad de 256 dimensiones entrenada con InfoNCE y una proyección lineal de cinco salidas entrenada con MSE contra las etiquetas del teacher. Una regularización fuerza ortogonalidad entre direcciones de la cabeza OCEAN; por ello la ortogonalidad es una propiedad impuesta por el objetivo, no un fenómeno descubierto espontáneamente. Tampoco se fuerza ni mide ortogonalidad entre personalidad, identidad y circuitos de razonamiento. Sobre 1.000 ejemplos de validación, la mejor MSE es 0,0113 y la final 0,0118. El texto convierte 0,0113 en «≈99 % accuracy», pero MSE y exactitud no son métricas intercambiables, y no ofrece MAE, R², correlaciones, calibración, baselines, intervalos ni resultados por rasgo o personaje. Además, la contribución anuncia MSE < 0,01, que la tabla no alcanza. Un t-SNE de los mismos 1.000 embeddings, coloreado solo por la etiqueta de Openness del teacher, muestra dos agrupaciones visuales; t-SNE no prueba ortogonalidad, continuidad global, separación de los cinco rasgos ni independencia respecto al razonamiento. Para steering, el vector es la diferencia normalizada entre la media de una persona objetivo y una media neutral, sumada al residual con intensidad α. Aunque el método formula intervención en la capa L−1, la ablación recorre capas 10–20 y declara un óptimo 14–16 con boost 6–8. Las métricas «Villainy» y «Sanity» de los heatmaps no se definen, no tienen muestra, evaluador, repeticiones o incertidumbre, y el caption confunde color oscuro con mayor adherencia aunque la leyenda de Villainy usa tonos rojos para valores altos. No hay generaciones de ejemplo, comparación con prompting/SFT/LoRA, prueba en perfiles nuevos ni evaluación humana. Sobre todo, no se ejecuta ningún benchmark de razonamiento o inteligencia; por tanto, las afirmaciones de «alineamiento sin impuesto», «inteligencia original mantenida» y subespacios distintos de razonamiento no están probadas. El estudio es una demostración preliminar en un único modelo de 0,5B con etiquetas de otro LLM. Aporta una arquitectura posible y un error de regresión interno, pero no valida personalidad psicológica, control determinista, preservación de capacidades ni seguridad. El «Safety Interceptor» y la sustracción de vectores maliciosos son trabajo futuro, no resultados.

Research question

Can a partially fine-tuned encoder learn from OCEAN labels generated by a teacher, separate stylistic representations, and enable steering by activation differences without degrading the general capabilities of the model?

Method

SoulBench dynamically concatenates three sentences of the same character and uses OCEAN labels generated by Doubao-Seed-1.6. In Qwen2.5-0.5B-Instruct, layers 0–19 are frozen and layers 20–23 are fine-tuned. An identity head uses InfoNCE; an OCEAN head uses MSE and orthogonal regularization. Validation reports MSE over 1,000 examples, a t-SNE of Openness, and a grid of layers 10–20 and intensities 2–15 with undefined metrics of Villainy/Sanity.

Sample: A single 0.5B model. Validation contains 1,000 chunks of three character sentences; the training size, number of characters, and trait distribution are not reported. The steering grid covers six layers and seven boosts, but it is not stated how many generations support each cell.

Findings

  • The OCEAN head obtains minimum MSE 0.0113 and final MSE 0.0118 against labels generated by Doubao; this is not a comparison with human psychological scores.
  • The t-SNE colored by Openness shows two clusters, but it does not quantify separation and does not evaluate the other four traits or orthogonality.
  • The heatmap places the best declared trade-off between Villainy and Sanity at layers 14–16 and boosts 6–8, although both metrics and their measurement protocol are not defined.
  • Orthogonality between OCEAN directions is explicitly incorporated as a training penalty; the result does not demonstrate that it was present in the base model.
  • The upper layers 20–23 are fine-tuned, in contradiction with the description of a frozen backbone or intervention without weight modification.
  • No results of MMLU, reasoning, syntax, human coherence, stability, or safety are reported to support preservation of intelligence.

Limitations

  • SoulBench is not auditable: provenance, composition, licenses, number of characters, training size, label distribution, duplicates, and definition of neutrality are missing.
  • The OCEAN labels come from a single LLM teacher and character profiles, with no questionnaire, humans, reliability, convergent validity, or inter-rater agreement; "psychological ground truth" is an overstatement.
  • MSE 0.0113 does not equate to 99% accuracy. Without baselines and per-dimension metrics, utility cannot be judged or compared with a trivial mean.
  • The regularization imposes orthogonality within the psychometric head and does not explicitly separate that head from identity or reasoning; the conclusion of independent subspaces does not follow from the objective.
  • t-SNE distorts distances and only colors Openness; a 2D visualization does not confirm global geometry, causality, or continuity of five traits.
  • Villainy and Sanity lack definition, evaluator, scale, examples, repetitions, and error; the heatmap does not constitute a reproducible ablation.
  • Only Qwen2.5-0.5B is studied and its last layers are fine-tuned. Transfer to 7B/70B, new profiles, languages, tasks, or models not trained with SoulBench is not tested.
  • There is no capability benchmark, prompting/SFT/LoRA baseline, human evaluation, or harm analysis; the safety mechanism is purely hypothetical.

What the study does not establish

  • It does not demonstrate that personality exists naturally as a linear or orthogonal subspace in Qwen or in LLMs in general; the geometry is trained with an explicit constraint.
  • It does not demonstrate independence between personality and intelligence or absence of alignment tax, because it does not measure reasoning before and after.
  • It does not validate OCEAN labels, Villainy, Sanity, or characters as human psychological constructs.
  • It does not prove deterministic or stable control: it does not report seeds, variance, success rate, examples, or longitudinal evaluation.
  • It does not demonstrate intervention without weight changes, since the encoder fine-tunes the four upper layers of the backbone.
  • It does not demonstrate that the Safety Interceptor detects or eliminates malicious intent; it is a future proposal and potentially risky.

Traceability

Scope: Full text

Version: arXiv:2512.07092v1; preprint under review

Consulted source: https://arxiv.org/pdf/2512.07092

Review: Codex editorial review, 2026-07-14

Approval: Codex fidelity pass, 2026-07-14

English translation: approved, 2026-07-18

Models evaluated

  • Qwen2.5-0.5B-Instruct
  • Doubao-Seed-1.6 (teacher for OCEAN labels)

Instruments and metrics

  • Teacher-generated OCEAN labels in [0,1]
  • InfoNCE contrastive loss
  • Mean squared error psychometric regression
  • Orthogonality regularization
  • t-SNE visualization
  • Undefined Villainy and Sanity heatmap scores

Data used

  • SoulBench character-sentence corpus (sources and size not reported)

Evidence and location

  • Hypothesis, preservation promises, and contributions: arXiv v1, pp. 1–2, Abstract and Introduction
  • SoulBench and origin of the OCEAN labels: arXiv v1, p. 3, section 2.1
  • Actually fine-tuned layers and two-head architecture: arXiv v1, pp. 3–4, sections 2.2–2.3 and Figure 1
  • Orthogonality imposed by the loss function: arXiv v1, p. 4, equation 6
  • Steering vector and inconsistency of intervention layer: arXiv v1, pp. 5–7, section 2.4 and Figure 3
  • MSE, sample, and invalid conversion to accuracy: arXiv v1, p. 5, sections 3.1–3.2 and Table 1
  • Real scope of t-SNE and heatmaps: arXiv v1, pp. 6–7, Figures 2–3
  • Absence of proof at scale and future Safety Interceptor: arXiv v1, pp. 7–8, sections 4.3–5