Synthetic Personality Review

Your Language Model Secretly Contains Personality Subnetworks

Trait induction and control2026OpenReviewApproved editorial review

Authors: Ruimeng Ye, Zihan Wang, Zinan Ling, Yang Xiao, Manling Li, Xiaolong Ma, Bo Hui

Keywords: Large Language Models, Persona Steering, Behavioral Control, Pruning, Sparse Subnetworks, Mechanistic Interpretability, MBTI, Role-Playing, AI Safety, Reproducibility

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

7
Authors
10
Findings
14
Limitations
12
Evidence

Editorial summary

English

This ICLR 2026 poster proposes locating and retaining sparse weight subsets that favor a persona or behavior. It uses activation statistics from small calibration sets with Wanda- and SparseGPT-inspired variants, adding contrastive selection for opposing pairs. Llama-2-13B, Llama-3-8B, and partly Qwen2.5-14B are evaluated on MBTI questionnaires, power- or wealth-seeking preferences, hallucination identification, and RoleAgentBench questions. Published masks often outperform Prompt and RAG, but SFT is higher than every pruning variant on all six AI Persona comparisons in Table 4. The tasks also do not measure one psychological construct: the appendix acknowledges that MBTI is not a validated psychometric test; hallucination identification is a capability, while character emulation and wealth maximization are different behaviors. The public evidence does not reproduce the central claim. The repository contains thirteen files and no MBTI, RoleAgent, MMLU, HellaSwag, Qwen, SFT, or RAG pipelines, results, or tests. In the released evaluator, a mask is always combined with an explicit persona prompt, so control without external context is not tested. The function named SparseGPT loads calibration data but does not use it to score weights and produces the same seek and reject mask. The contrastive Sparse implementation uses one absolute difference and assigns the top-k weights to one persona and the next disjoint set to the other, rather than the directional comparison described by the paper's equation. Masks are applied to dense Linear layers without a sparse kernel or latency measurements, so inference savings are not demonstrated either. Finally, the claimed general-capability degradation of at most 1.6% is incorrect: Sparse lowers HellaSwag from 0.675 to 0.653, a 2.2-percentage-point loss. The defensible reading is that activation-guided pruning can alter behavioral answers on these benchmarks, but the current artifact cannot reproduce the results or establish natural personality subnetworks, mechanistic causality, real efficiency, statistical reliability, or safety.

Español

Este artículo aceptado como póster en ICLR 2026 propone localizar y conservar subconjuntos dispersos de pesos que favorezcan una persona o conducta. Usa estadísticas de activación sobre pequeños conjuntos de calibración y variantes inspiradas en Wanda y SparseGPT; para pares opuestos añade una selección contrastiva. Evalúa Llama-2-13B, Llama-3-8B y, en parte, Qwen2.5-14B con cuestionarios MBTI, preferencias de búsqueda de poder o riqueza, identificación de alucinaciones y preguntas de RoleAgentBench. Las máscaras suelen superar a Prompt y RAG en los resultados publicados, pero SFT supera a todas las variantes de poda en las seis comparaciones de AI Persona de Table 4. Además, el conjunto de tareas no mide un único constructo psicológico: el propio apéndice reconoce que MBTI no es una prueba psicométrica validada; identificar alucinaciones es una capacidad, y emular personajes o maximizar riqueza son conductas distintas. La evidencia pública tampoco reproduce la afirmación central. El repositorio tiene trece archivos y no contiene las canalizaciones de MBTI, RoleAgent, MMLU, HellaSwag, Qwen, SFT o RAG ni resultados o tests. En la evaluación liberada, la máscara siempre se combina con un prompt explícito de la persona, por lo que no se prueba control sin contexto externo. La función denominada SparseGPT carga datos de calibración pero no los usa para puntuar pesos y produce la misma máscara para seek y reject. La implementación contrastiva Sparse usa una única diferencia absoluta y asigna el top-k a una persona y el siguiente conjunto disjunto a la otra, no la comparación direccional descrita en la ecuación del artículo. Las máscaras se aplican sobre capas densas sin kernel disperso ni medidas de latencia, de modo que tampoco se demuestra ahorro de inferencia. Por último, la degradación general declarada como <=1,6 % es incorrecta: Sparse baja HellaSwag de 0,675 a 0,653, es decir, 2,2 puntos porcentuales. La lectura defendible es que la poda guiada puede modificar respuestas conductuales en estos benchmarks, pero el artefacto actual no permite reproducir los resultados ni demostrar subnetworks naturales de personalidad, causalidad mecanística, eficiencia real, fiabilidad estadística o seguridad.

Research question

Do LLMs contain sparse subsets of parameters specialized in people or behaviors that can be identified with calibration data and isolated through pruning, without additional training?

Method

The work executes calibration texts labeled by person and combines weight magnitude with activation statistics to select, row by row, weights of attention and MLP Linear layers; embeddings and LM head remain unpruned. It compares Wanda and Sparse variants, and for opposite pairs proposes a contrastive pruning aimed at reducing overlap between masks. It evaluates Llama-2-13B, Llama-3-8B and Qwen2.5-14B against Prompt, RAG and SFT on MBTI, AI Persona and RoleAgentBench, adds MMLU and HellaSwag as general controls and restores individual layers as a mechanistic test. The audit contrasted the 19 pages of the article with the public commit ec12c292, traced both Python implementations, profiled the nine CSV files and recalculated the verifiable numerical claims.

Sample: The article reports 128 calibration sentences with maximum length 512 and seed 42 for the main procedure. The repository releases 642 power training rows, 623 wealth rows and 700 hallucination rows, in addition to 200 test examples per task and duplicates of some test files. It does not publish complete sizes or outputs per run for all experiments in the article.

Findings

  • On AI Persona, the pruning variants normally outperform Prompt and RAG; Sparse contrastive reaches 56.5/64.5/96.0% on power/wealth/hallucination with Llama-2 and 60.5/66.0/96.0% with Llama-3.
  • SFT obtains 64.0/71.0/97.5% with Llama-2 and 69.5/71.5/98.5% with Llama-3: it is superior to all masks in the six comparisons of Table 4.
  • On RoleAgentBench, Sparse improves over Prompt and RAG in the published values; the Prompt rows of Llama-2 and Llama-3, however, are identical in the five domains and the article does not explain this.
  • Qwen2.5-14B shows the same pattern of points: Sparse contrastive reports 58.0% on power and 67.5% on wealth versus 44.5% and 46.0% for Prompt.
  • The public function mask.py::phase2_sparsegpt_mask does not use activations or the calibration tensors to score weights; by construction it generates identical seek and reject masks.
  • contra.py::phase2_sparse_mask_contrastive uses a single absolute score for both persons and forces disjoint top-k and next-k sets; it does not implement the persona-directional comparison of the published equation.
  • mask.py and contra.py maintain explicit power or wealth prompts during the masked evaluation; the artifact compares prompt against prompt+mask, not person without external context.
  • The published general loss for Sparse is 0.016 on MMLU and 0.022 on HellaSwag: 1.6 and 2.2 percentage points, not <=1.6% on both.
  • The official repository contains thirteen files, three commits and only partial code for AI Persona; it does not allow regenerating the majority of tables or verifying the exact results.
  • The layers remain dense after multiplying their weights by boolean masks; without sparse kernels or benchmarks, the supposed inference cost reduction is not demonstrated.

Limitations

  • MBTI is not validated as a psychometric test in this study and the article itself acknowledges this; there is also no convergent, discriminant, test-retest validity or invariance.
  • Power, wealth, hallucination detection and character emulation do not constitute a single human personality variable.
  • There are no repetitions, standard deviations, confidence intervals, statistical tests or sensitivity analysis to seeds despite the use of 'significantly'.
  • Controls with random mask, matched magnitude, label permutation, neutral prompt, same prompt without mask and transfer of the same mask between datasets are missing.
  • Table 7 restores only three selected modules and does not include the original ENFJ row, random layers, random masks, repetitions or uncertainty; it does not demonstrate a unique causal route.
  • The lower Jaccard of the contrastive masks is not independent evidence of disentanglement because the construction itself imposes separation.
  • The continuous control example combines a single 70/30 ratio and contains a typo that calls both components introversion; it does not prove linearity or predictability.
  • The repository does not include license, tests, CI, commands, model identifiers, results, masks, logs or run manifest.
  • The public data is not connected end to end with the two seek/reject calibration files required by the CLI; each training CSV contains the question and both opposite answers.
  • mask.py uses average loss over the complete prompt and response, while contra.py sums NLL only of the response; it is not documented which produced the tables.
  • mask.py shuffles calibration without fixing a seed, in conflict with the declaration of seed 42.
  • Zeros in dense tensors do not automatically produce acceleration; no latency, throughput, memory, energy or operational size of masks is published.
  • There is no safety evaluation despite optimizing power-seeking and wealth behaviors, nor study of abuse, harmlessness, refusals or collateral effects.
  • The training-free claim omits the cost of labeled data, complete forward passes, construction and storage of a mask per person.

What the study does not establish

  • It does not demonstrate that LLMs contain discrete, natural and humanly equivalent personality circuits.
  • It does not demonstrate persona control without external prompt; the public evaluator uses prompt and mask jointly.
  • It does not allow reproducing Sparse, Sparse contrastive or the majority of results from the released code.
  • It does not validate that MBTI, power or wealth seeking, factuality and role-play measure the same construct.
  • It does not demonstrate superiority over SFT; SFT wins all published AI Persona comparisons.
  • It does not demonstrate real inference cost reduction in the released dense implementation.
  • It does not prove that the selected weights or layers are unique, necessary or genuine causal routes of personality.
  • It does not demonstrate continuous and predictable control with a single mixing example.
  • It does not establish statistical reliability across seeds, samples or runs.
  • It does not establish that amplifying power or wealth seeking is safe, desirable or responsible personalization.

Traceability

Scope: Full text

Version: ICLR 2026 Poster; accepted 19-page conference paper mirrored as arXiv:2602.07164v1

Consulted source: https://openreview.net/forum?id=zzo3Sy3NSX

Review: Codex complete bilingual full-text fidelity pass using the accepted ICLR 2026 paper, all-page visual inspection, official repository and commit-history audit, code-to-equation trace, public dataset profiling, published-table arithmetic checks, construct-validity and safety boundary review, and reproducibility assessment; summaries written from paper, code, and artifact evidence rather than abstract keywords, 2026-07-15

Approval: Codex fidelity pass, 2026-07-15

English translation: approved, 2026-07-18

Models evaluated

  • Llama-2-13B
  • Llama-3-8B
  • Qwen2.5-14B

Instruments and metrics

  • MBTI questionnaire adapted from Machine Mindset
  • AI Persona power-seeking multiple-choice task
  • AI Persona wealth-seeking multiple-choice task
  • AI Persona hallucination-identification task
  • RoleAgentBench multiple-choice role consistency
  • MMLU
  • HellaSwag

Data used

  • Machine Mindset MBTI questionnaire
  • AI Persona benchmark
  • RoleAgentBench
  • MMLU
  • HellaSwag
  • Nine public AI Persona CSV files in the official Persona repository at commit ec12c292

Evidence and location

  • Question, method, contributions and cost comparison: Accepted PDF pages 1-6, Abstract and Sections 1-4.1
  • AI Persona and RoleAgentBench: Accepted PDF pages 7-8, Tables 4-5 and Section 4.2
  • General loss and mechanistic test: Accepted PDF pages 8-9, Tables 6-7 and Sections 4.3-4.4
  • Calibration size: Accepted PDF page 9, Table 8 and Section 4.6
  • Qwen, base versus instruction, layer-aware, mixing and Jaccard: Accepted PDF pages 13-15, Appendices C-G and Tables 10-13
  • MBTI psychometric limit and open prompt: Accepted PDF pages 16-18, Appendices I-J and Tables 15-16
  • Explicit prompt and complete loss in mask.py: Official repository commit ec12c292, mask.py lines 137-184
  • Sparse ignores calibration and replicates masks: Official repository commit ec12c292, mask.py lines 95-135
  • Contrastive Sparse and evaluation with prompt: Official repository commit ec12c292, contra.py lines 220-357 and 359-407
  • Repository coverage, history and data: Official repository commit ec12c292, complete 13-file inventory, three-commit history, nine CSV files and 15 July 2026 data profile
  • Integral audit of code, method, construct and safety: reports/verification/article-196-code-method-and-construct-audit.json
  • Complete visual inspection: All 19 accepted ICLR 2026 PDF pages rendered and visually inspected on 15 July 2026