Source-linked AI summary
When Personality Meets Quantization: A Layer-wise MBTI Analysis of Quantized LLMs
Yao Fu, Lijia Huang, Xiaomin Li, Runchao Li, Yu Yin, Kenneth A. Loparo
TL;DR
Existing MBTI evaluations have focused mainly on full-precision LLMs and final outputs, leaving quantized models and internal personality dynamics underexplored. This paper evaluates open-source models across precisions and prompting conditions, analyzes layer-wise uncertainty, and introduces UALD for inference-time decoding. It finds that personality is layer-dependent and sensitive to quantization, prompting, and decoding, with ENFJ remaining dominant while 2-bit quantization and decoding can reduce consistency or shift personality.
Problem
Existing MBTI studies primarily examine full-precision models and final outputs, while quantized LLM personality remains underexplored.
Method
The paper systematically evaluates open-source LLMs across multiple precisions and prompting conditions, analyzes layer-wise entropy and confidence gaps, and introduces UALD for decoding analysis.
Results
ENFJ remains dominant across model families and quantization settings, while 4-bit behavior is largely stable and 2-bit quantization reduces controllability and cross-precision consistency.
Takeaways & Limitations
Personality decisions emerge through upper-layer sharpening after earlier ambiguity, and personality-aligned prompting improves robustness under some perturbations.
Takeaways & Limitations
The study is limited to selected model scales and quantization methods, single-turn MBTI prompts, probability-based analyses without causal interpretation, and MBTI itself.
Abstract
from arXiv · showhide
Personality is increasingly important in large language models (LLMs), as it shapes users' trust, engagement, and emotional experiences. While the Myers--Briggs Type Indicator (MBTI) has emerged as a common framework for assessing LLMs' personality, existing studies focus primarily on full-precision models and evaluate only final outputs. They overlook the widespread deployment of quantized LLMs requiring low memory footprints, whose personality traits remain underexplored. In this work, we present a systematic MBTI analysis of open-source LLMs across multiple precisions, including mainstream 4-bit methods (GPTQ, AWQ) and extreme 2-bit settings (AQLM variants). Beyond output-level evaluation, we examine how personality emerges across layers through option-level entropy and confidence-gap dynamics, and introduce Uncertainty-Amplified Layer Decoding (UALD) to study decoding-induced personality drift at inference time. Our results reveal a key insight: LLMs' personality is not a static property, but an emergent, layer-dependent decision process sensitive to quantization, prompting, and decoding. Specifically, we find that (1) ENFJ remains dominant across model families and precisions; (2) 4-bit quantization largely preserves coarse personality structure, while 2-bit quantization disrupts fine-grained prompt consistency and cross-precision agreement; (3) personality decisions emerges in upper layers, following substantial ambiguity in early layers; and (4) inference decoding can shift personality, while personality-aligned conditioning improves robustness. These findings provide a new perspective on the behavioral reliability of quantized LLMs and highlight the importance of considering internal dynamics and inference strategies in personality-sensitive chatbot applications.
1 Introduction
This paper investigates whether LLM personality remains stable after quantization and under different prompting and decoding conditions. It evaluates quantized open-source models, examines layer-wise decision dynamics, and studies decoding-induced personality drift.
- Motivation: Existing MBTI studies mainly assess full-precision LLMs and final outputs, leaving the personality of compressed models underexplored.Quantization is widely used to reduce deployment memory and inference cost.
- Approach: The paper evaluates open-source LLMs across multiple precision regimes, including mainstream 4-bit and extreme 2-bit quantization settings.The pipeline covers unconditional and personality-conditional prompting.
- Approach: The analysis extends beyond final outputs by tracking layer-wise entropy and confidence-gap dynamics during MBTI choice prediction.These measures characterize how uncertainty develops into personality commitments across layers.
- Approach: UALD studies how inference-time decoding can amplify intermediate-layer uncertainty and induce personality drift.It is designed for discrete personality choices rather than factual correctness.
- Key insight: The study frames personality as an emergent, layer-dependent process sensitive to quantization, prompting, and decoding rather than a static model property.This distinction matters because MBTI items lack externally verifiable ground truth.
2 Related Work
Related work covers quantization methods and broader behavioral evaluations of compressed LLMs, while prior personality research has largely examined uncompressed models. This paper positions systematic personality assessment as a distinct evaluation direction for quantized LLMs.
- LLM Quantization: Quantization reduces storage and computational requirements by mapping high-precision parameters to lower-precision representations.Existing approaches are commonly divided into post-training quantization and quantization-aware training.
- LLM Quantization: Post-training quantization is generally more efficient, but it often yields lower performance than quantization-aware training.
- Compressed LLM Evaluation: Compressed LLMs have been evaluated for safety, toxicity, bias, trustworthiness, long-context inputs, and long-form generation.
- Research Gap: Unlike previous compressed-LLM studies, this work focuses on systematic personality assessment of quantized LLMs.
- LLM Personality Research: Prior LLM personality research has examined intrinsic traits, prompt-based personality instillation, and related personality directions.
3 Experimental Settings for MBTI Analysis
The experimental framework combines MBTI evaluation with representative open-source models, multiple quantization methods, two prompting paradigms, deterministic choice prediction, and inference-time decoding analysis. It examines both intrinsic tendencies and controllability under externally specified personalities.
- Models: The study evaluates representative LLaMA, Mistral, and Qwen models selected for public availability, comparability, performance, and practitioner adoption.
- Quantization: Quantization experiments cover widely adopted GPTQ and AWQ methods alongside extreme 2-bit AQLM and PV-tuned AQLM.
- MBTI Fundamentals: MBTI measures four binary dichotomies and defines 16 personality types; the study uses a standardized 60-item questionnaire.
- Prompting: Unconditional prompting probes default personality tendencies, whereas personality-conditional prompting tests shifts under externally imposed MBTI constraints.
- Decoding: The framework reformulates MBTI assessment as single-token choice prediction to reduce sampling variability from multi-token generation.
- Inference Decoding: UALD compares mature-layer predictions with intermediate-layer candidates and uses an evolution scale λ to amplify intermediate uncertainty.The tested values are λ ∈{5, 10, 15, 20, 25, 30, 35, 40}.
4 MBTI Personality Analysis
MBTI analysis finds a dominant ENFJ attribution across model families and quantization settings, while layer-wise evidence shows personality decisions sharpening late after early ambiguity. Prompting, model size, quantization level, and decoding influence consistency and drift, with extreme 2-bit settings showing greater instability.
- Personality Assessment Based on Outputs: ENFJ remains the dominant MBTI type across model families and quantization settings.The authors associate this prevalence with empathetic, warm, supportive, and guidance-oriented responses.
- Layer-wise Personality Formation: Early and intermediate layers show high entropy and small top-1/top-2 probability gaps, indicating unresolved preference structures.This ambiguity persists approximately through layers 1 to 21 across model variants.
- Layer-wise Personality Formation: Upper layers show declining entropy and increasing confidence gaps, producing increasingly decisive MBTI predictions.This commitment emerges approximately in layers 22 to 32; full-precision and 4-bit models sharpen smoothly, whereas 2-bit models sharpen later or less stably.
- Personality-Conditional Prompting: Personality conditioning reveals asymmetric prompt consistency: Extraversion and Intuition remain stable across precisions, often overriding opposing conditioned traits.The findings characterize configurations such as ENFJ as hierarchical latent priors that are difficult to manipulate through prompts.
- Personality-Conditional Prompting: Larger models are more robust to prompt variation, while smaller models are more sensitive and quantization reduces robustness.GPTQ-INT4 and AWQ-INT4 largely preserve smaller-model agreement with FP16, whereas extreme 2-bit models often show unstable predictions and poor conditional prompt following.
- Decoding-Induced Personality Drift: Uncertainty-Amplified Layer Decoding can shift inferred personality, while conditioning aligned with the intrinsic ENFJ prior improves robustness.Under unconditional prompting, FP16 drifts across multiple types as evolution scale increases; GPTQ and AWQ show heterogeneous sensitivity, while Mistral-7B remains largely stable.
5 Conclusions
The analysis finds that personality remains largely stable under 4-bit compression but becomes less controllable and consistent at 2 bits. Personality decisions emerge progressively across layers, while inference decoding can induce personality drift.
- 4-bit quantization largely preserves personality behavior, whereas 2-bit quantization degrades controllability and cross-precision consistency.
- Personality decisions develop from high-entropy intermediate states into sharper final commitments across layers.
- Uncertainty-Amplified Layer Decoding can induce personality drift during inference.
Limitations
The study is limited in model and quantization coverage, dialogue setting, interpretive methodology, and personality framework generality.
- The evaluation excludes some model scales, including LLaMA3.1-405B and Qwen3-235B, and quantization methods such as QAT and PEFT.
- The study evaluates single-turn MBTI-style prompts that may not represent multi-turn dialogue.
- The analyses rely on probability-based metrics without providing causal interpretation.
- The results are specific to MBTI and may not generalize to other personality frameworks.
Ethical Considerations
The paper frames personality labels as conditional behavioral outputs rather than fixed human-like identities and highlights risks in user-facing deployment. It recommends transparency, cautious use, and behavioral monitoring.
- Personality labels may be over-interpreted anthropomorphically as stable human-like traits despite sensitivity to quantization, prompting, and decoding.
- Personality conditioning could increase persuasive influence or emotional dependence in education, counseling, and decision support.
- Transparent disclosure of persona-conditioning settings, cautious deployment, and monitoring for behavioral drift are recommended safeguards.
- The evaluation uses unconditional and personality-conditioned assessment prompts across quantized settings.
B MBTI Personality Descriptions and Traits
The appendix describes MBTI-associated trait profiles through recurring patterns involving organization, responsibility, intuition, independence, practicality, and interpersonal consideration.
- Several profiles emphasize thoroughness, dependability, loyalty, organization, practicality, and responsibility.
- Across profiles, stressors include disorder, indecision, conflict, noise, criticism, disrupted routines, and insufficient preparation.
- Some profiles emphasize meaning, connection, compassion, insight, idealism, and service to the common good.
- Other profiles emphasize original ideas, pattern recognition, strategic thinking, independence, competence, and intellectual challenge.
- Practical problem-solving profiles emphasize flexibility, rapid response, analysis of cause and effect, efficiency, and autonomy.
C MBTI Scoring from Seven-Option Responses
The assessment converts seven-option responses into signed scores, aggregates them within four MBTI dichotomies, and concatenates the selected poles into a final type.
- Each model answers 60 MBTI-style questions using a seven-point response scale from Agree to Disagree.
- Each question belongs to one pole of an MBTI dichotomy: {E, I}, {N, S}, {F, T}, or {J, P}.
- The seven response options are mapped to scores from +3 for A through −3 for G, with neutrality assigned zero.This preserves agreement direction, disagreement direction, and response strength.
- For each dichotomy, keyed question scores are aggregated into a dichotomy score, whose sign determines the predicted pole.
- The final MBTI type concatenates the four predicted poles, producing ENFJ when the predictions are E, N, F, and J.
D Layer-wise Entropy and Confidence Gap over Seven Options
The paper tracks seven-option probabilities across transformer layers using entropy and top-two confidence gaps. Their joint trajectory characterizes movement from ambiguous intermediate representations toward stable personality decisions.
- Option-level logits extracted at each layer are converted with softmax into a probability vector over the seven response choices.
- Layer-wise Shannon entropy measures uncertainty: high values indicate probability spread across options, while low values indicate more decisive behavior.
- The average entropy summarizes uncertainty across all N = 60 questions at each layer.
- The confidence gap measures how strongly the highest-probability option exceeds the second-highest option.Larger gaps indicate clearer preference, whereas smaller gaps indicate that alternatives remain comparably plausible.
- Decreasing entropy and increasing confidence gaps together describe a transition from diffuse uncertainty to concentrated, stable personality judgments.
E Details of UALD
UALD probes how intermediate-layer ambiguity influences MBTI decoding by comparing premature-layer and mature-layer option distributions. Its evolution scale provides a controllable way to shift decoding toward earlier-layer structure.
- UALD collapses vocabulary logits into option-level logits over the seven valid choices before computing probability distributions.
- For each candidate premature layer, UALD computes Jensen–Shannon divergence relative to the mature layer and selects a premature layer for decoding.
- The final UALD-adjusted logits combine mature-layer and selected premature-layer information, with λ controlling the evolution scale.
- Increasing λ shifts the decoding geometry toward earlier-layer probability structure, interpolating between mature certainty and intermediate-layer ambiguity.
F Download Links of Models
This section provides model download links and presents layer-wise diagnostic figures and conditional-consistency tables across model families, quantization variants, and prompting conditions.
- Table 6 lists download links for all LLMs involved in the experiments.
- Figures 4–8 compare Original FP16, GPTQ INT4, AWQ INT4, and Extreme INT2 variants across layers for five model families.Each figure reports layer-wise decisional entropy and confidence gap.
- Tables 7–14 report conditional personality consistency and switch statistics across quantization levels and personality-conditioned prompts.
- Tables 15–17 continue switch-statistics reporting for Mistral-Small-24B-Instruct-2501, Qwen2.5-14B-Instruct, and Qwen2.5-72B-Instruct.
- Figures 9 and 10 show early-layer uncertainty followed by progressively sharper decisions in deeper layers for LLaMA3.1-70B-Instruct and Mistral-7B-Instruct-v0.3.
- Figure 11 shows lower entropy and larger confidence margins in deeper Mistral-Small-24B layers, while stronger quantization can disrupt late-layer consolidation.