💬 NLP

When Personality Meets Quantization: A Layer-wise MBTI Analysis of Quantized LLMs

This paper presents a systematic layer-wise MBTI analysis of quantized LLMs, revealing that personality is an emergent, layer-dependent process sensitive to quantization levels and decoding strategies, with 4-bit models largely preserving personality structures while 2-bit models disrupt fine-grained consistency.

Original authors: Yao Fu, Lijia Huang, Xiaomin Li, Runchao Li, Yu Yin, Kenneth A. Loparo

Published 2026-08-27

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yao Fu, Lijia Huang, Xiaomin Li, Runchao Li, Yu Yin, Kenneth A. Loparo

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the rapidly evolving world of artificial intelligence, large language models have moved beyond simple text completion to become interactive companions, offering advice, comfort, and conversation. As these digital entities become more integrated into daily life, particularly for younger users who increasingly turn to them for serious discussions, the way they sound and behave has taken on new importance. People do not just want accurate answers; they want interactions that feel trustworthy, warm, and emotionally resonant. To understand this behavior, researchers often look to the Myers-Briggs Type Indicator, a well-known framework that categorizes human personalities into sixteen distinct types based on how individuals process information and make decisions. While previous studies have examined whether these AI models possess a consistent personality, they have largely focused on the most powerful, full-sized versions of the software. However, in the real world, these massive models are often compressed to run on smaller devices or to save energy, a process that reduces the amount of memory they require. It remains unclear whether this compression changes the personality of the AI, turning a warm, empathetic companion into something colder or more erratic.

A team of researchers from Case Western Reserve University, Northeastern University, and Microsoft Research set out to investigate this question by treating personality not as a fixed label, but as a dynamic process that unfolds inside the model. They tested several popular open-source AI models, including versions of LLaMA, Mistral, and Qwen, running them in their original high-precision form and then in compressed versions that use significantly less memory. Some of these compressed versions were reduced to four bits of precision, a common standard for efficient deployment, while others were pushed to an extreme two-bit setting to see how far the models could be squeezed before their behavior broke down. The researchers did not simply ask the models to answer a personality questionnaire and record the final result. Instead, they developed a method to watch how the models made their choices layer by layer as they processed each question, observing how their internal confidence shifted from uncertainty to a firm decision. They also introduced a technique to deliberately introduce uncertainty during the reading process, allowing them to see if the models' personalities would drift or remain stable when the path to an answer was made more difficult.

The study revealed that the personality of these artificial intelligence systems is not a static trait but an emergent property that changes depending on how the model is configured and how it is asked to respond. Across all the different model families and compression levels tested, a specific personality type known as ENFJ consistently emerged as the dominant character. This type is characterized by being outgoing, intuitive, feeling-oriented, and organized, suggesting that the training processes used to make these models helpful and harmless naturally steer them toward a warm, supportive, and guidance-focused persona. The researchers found that compressing the models to four bits generally preserved this broad personality structure, meaning the AI still sounded like the same helpful assistant. However, when the models were compressed to the extreme two-bit level, the fine-grained details of their personality began to fracture. While the general warmth remained, the models struggled to maintain consistency when asked to adopt specific, different personality traits, and their ability to agree with their own un-compressed versions on nuanced questions diminished.

Perhaps the most revealing part of the research was the observation of how these personality decisions actually happen inside the machine. When the models were processing a question, the early layers of their internal structure showed a high degree of confusion, with the probability of choosing any specific answer spread out thinly across many options. It was only as the information moved through the deeper layers of the network that the model began to narrow its focus, eventually settling on a single, decisive answer. This progression from ambiguity to certainty mirrors the way humans often arrive at a self-assessment, moving from a vague sense of preference to a clear conclusion. The researchers discovered that this decision-making process is sensitive to the method used to generate the final answer. When they manipulated the decoding process to amplify the uncertainty from the earlier, confused layers, the models' predicted personality types would shift, sometimes changing entirely. This indicates that the personality attributed to an AI is not just a fixed output but a result of the specific path the model takes to reach a conclusion.

The study also explored whether these models could be guided to act like different personality types if explicitly instructed to do so. The results showed a strong resistance to change in certain core traits. Even when told to act as a reserved or logical type, the models frequently reverted to their default warm and intuitive nature, suggesting that their underlying personality is deeply embedded in their training rather than being a superficial layer that can be easily peeled away. Larger models proved more robust in maintaining their core identity under these instructions, while smaller models were more easily swayed. However, the extreme compression to two bits disrupted this stability, causing the models to behave unpredictably and fail to follow even clear instructions about who they should be. The researchers concluded that while moderate compression is safe for maintaining the general character of an AI assistant, pushing these models to their absolute limits risks introducing instability in their behavior. This finding highlights that the reliability of a chatbot's personality depends not just on the model itself, but on the specific technical choices made in how it is stored and how it is asked to think.

Technical Summary: When Personality Meets Quantization: A Layer-wise MBTI Analysis of Quantized LLMs

Problem Statement
While Large Language Models (LLMs) are increasingly deployed as interactive assistants where personality shapes user trust and engagement, existing research on LLM personality—specifically using the Myers–Briggs Type Indicator (MBTI)—focuses almost exclusively on full-precision models and evaluates only final outputs. This overlooks the widespread practical deployment of quantized LLMs (e.g., 4-bit and 2-bit) used to reduce memory footprints and inference costs. Furthermore, current studies fail to address how personality judgments emerge across network layers or how inference-time decoding strategies influence personality attribution. The core research question is whether the personality traits attributed to LLMs reflect stable behavioral patterns or are sensitive to implementation factors such as quantization precision, prompting strategies, and decoding dynamics.

Methodology
The authors present a systematic MBTI evaluation pipeline covering open-source model families (LLaMA, Mistral, Qwen) across multiple precision regimes: Full Precision (FP16), mainstream 4-bit quantization (GPTQ, AWQ), and extreme 2-bit quantization (AQLM variants, PV-tuned AQLM).

Prompting Strategies: The study employs two paradigms:
1. Unconditional Prompting: Probes intrinsic personality dispositions without external conditioning.
2. Personality-Conditional Prompting: Tests controllability by instructing the model to adopt a specific MBTI type (e.g., "respond as an ENFJ").
Decoding Reformulation: To eliminate sampling stochasticity, MBTI assessment is reformulated as a single-token classification problem (predicting one of seven agreement options: A–G) using greedy decoding, rather than multi-token generation.
Layer-wise Analysis: The paper introduces metrics to track decision dynamics across layers:
- Decisional Entropy: Measures uncertainty in the probability distribution over response options.
- Confidence Gap: The probability difference between the top-1 and top-2 options, indicating decisiveness.
Uncertainty-Amplified Layer Decoding (UALD): A novel inference strategy inspired by DoLa but adapted for subjective tasks. UALD collapses vocabulary logits into option-specific scores and combines the mature (final) layer logits with those of a selected "premature" layer (the one with maximal Jensen–Shannon divergence from the mature layer). This is controlled by an evolution scale parameter ( $\lambda$ ) to study how amplifying intermediate uncertainty shifts personality attribution.

Key Contributions

First Systematic MBTI Evaluation for Quantized LLMs: The paper provides the first comprehensive analysis of personality traits across 4-bit and extreme 2-bit settings, covering both unconditional and personality-conditional prompting.
Layer-wise Personality Formation Analysis: Moving beyond output-level evaluation, the study characterizes how personality decisions evolve from high-entropy, ambiguous intermediate states in early layers to sharp, committed decisions in upper layers.
Decoding-Induced Drift Investigation: The authors introduce UALD to demonstrate that inference-time decoding strategies can induce personality drift, identifying conditions under which assessments become unstable.

Key Results

Dominant Personality Type: Across all model families and precision levels, ENFJ (Extraverted, Intuitive, Feeling, Judging) emerges as the dominant MBTI type. This suggests a systematic tendency for LLMs to produce empathetic, warm, and guidance-oriented responses, likely shaped by post-training processes like RLHF and DPO.
Quantization Impact:
- 4-bit Quantization: Largely preserves the coarse personality structure and cross-precision agreement with full-precision models.
- 2-bit Quantization: Disrupts fine-grained prompt consistency and cross-precision agreement. Extreme compression often leads to unstable personality predictions and a failure to follow conditional prompts, particularly in smaller models.
Layer-wise Dynamics: Personality decisions are not static but emerge progressively. Early and intermediate layers exhibit high entropy and small confidence gaps (ambiguity), while upper layers show a sharp drop in entropy and increased confidence gaps (decisional commitment). This pattern holds for FP16 and 4-bit models but is often delayed or unstable in 2-bit models.
Prompting and Robustness:
- Certain dimensions (Extraversion and Intuition) exhibit hierarchical stability, resisting prompt changes even when explicitly conditioned on opposing traits.
- Larger models (>70B) show robustness to prompt variations, while smaller models (<14B) are more sensitive. Quantization generally degrades this robustness, making models more susceptible to aligning with prompt conditioning or becoming unstable.
Decoding Drift: UALD experiments reveal that amplifying premature-layer logits can shift the inferred personality type. However, when the conditioning signal aligns with the model's intrinsic personality (e.g., ENFJ conditioning on an ENFJ-biased model), decisional commitment is reinforced, and the system remains robust to drift.

Significance and Claims
The paper claims that LLM personality is not a static property but an emergent, layer-dependent decision process sensitive to quantization, prompting, and decoding. The findings provide a new perspective on the behavioral reliability of quantized LLMs, highlighting that while moderate compression (4-bit) preserves personality traits, extreme compression (2-bit) can disrupt the fine-grained decision boundaries required for reliable attribution.

The authors emphasize that these results underscore the importance of considering internal dynamics and inference strategies in personality-sensitive chatbot applications. They caution against anthropomorphic over-interpretation of model-assigned personality labels, noting that such labels are conditional behavioral outputs rather than fixed identities. The study suggests that for high-stakes applications, transparent disclosure of persona-conditioning settings and monitoring protocols for behavioral drift are necessary, as compression and decoding choices can significantly alter model behavior even when standard benchmark performance remains stable.

More like this