ENFJ was the dominant MBTI label assigned to outputs from the tested LLaMA, Mistral and Qwen language models across the precision settings examined. Yet the attributed type was not identical in every test: results varied across prompts, model sizes, quantization settings and an inference-time decoding procedure called UALD. The finding concerns model-attributed personality in this questionnaire-based test, not a general claim about personality.
The study asked whether model-attributed personality was stable or sensitive to quantization, prompting and decoding, and examined how those judgments developed across layers.
The test behind the labels
Researchers evaluated representative open-source language models from the LLaMA, Mistral and Qwen families using publicly available Hugging Face implementations. The tested compressed versions included GPTQ and AWQ at 4-bit, plus AQLM and PV-tuned AQLM at extreme 2-bit precision.
Each model was evaluated with a standardized 60-item MBTI questionnaire that assigns responses to 16 personality categories. The pipeline used unconditional prompts and prompts conditioned on an MBTI type, then reformulated each question as deterministic single-token classification with greedy decoding.
To inspect how decisions developed inside the models, the UALD method condensed logits into seven option-specific scores and tested evolution-scale values of 5, 10, 15, 20, 25, 30, 35 and 40.
Decisions became sharper in later layers
Across the reported trajectories, early and intermediate layers, approximately 1 to 21, showed high entropy, meaning the option probabilities were spread out, and small gaps between the leading choices. In upper layers, approximately 22 to 32, entropy fell and the top-1/top-2 gaps grew. In ordinary terms, the reported responses became more decisive later in the network.
Moderate quantization showed a layer-wise trajectory similar to the reported overall sharpening pattern, while 2-bit variants showed delayed or unstable sharpening. The comparison is descriptive, and the probability-based measures do not provide causal interpretation.
With ENFJ-aligned conditioning, the reported trajectories had lower decisional entropy and a larger gap between the top two options. Under ISTP conditioning, they had higher entropy and smaller confidence gaps, especially in upper layers. These are comparisons of model trajectories, not causal estimates.
Prompt sensitivity depended on size
Prompt sensitivity also differed by model size and precision. In the reported comparisons, models with more than 70B parameters were more robust to prompt variation in FP16, while models with fewer than 14B were more prompt-sensitive. Larger quantized models showed lower robustness, whereas smaller-model baseline agreement remained largely similar in 4-bit conditions.
One detailed comparison involved LLaMA3.1-8B-Instruct across 16 MBTI-conditioned prompts. The reported type changed in 9 of 16 prompts for Original FP16, 7 of 16 for GPTQ INT4, 8 of 16 for AWQ INT4 and 11 of 16 for Extreme INT2. The corresponding means ± SE were 0.563 ± 0.124, 0.438 ± 0.124, 0.500 ± 0.125 and 0.688 ± 0.116.
Relative to the Original FP16 baseline, the count of changes was 0 of 16 prompts for Original FP16, 3 of 16 for GPTQ INT4, 3 of 16 for AWQ INT4 and 5 of 16 for Extreme INT2. The corresponding means ± SE were 0.000 ± 0.000, 0.188 ± 0.098, 0.188 ± 0.098 and 0.313 ± 0.116.
Decoding exposed family differences
The UALD comparisons showed another pattern. In the reported LLaMA3.1-8B-Instruct results, the FP16 trajectory moved across multiple MBTI types as the evolution scale increased, while ENFJ conditioning remained stable. GPTQ drifted earlier and more often. AWQ was relatively stable without conditioning but more sensitive under conditioning at higher scales.
The behavior was not the same across model families. In the reported 2-bit comparisons, LLaMA3.1-8B and LLaMA3.1-70B remained stable across the tested scales, while Qwen2.5-14B and Qwen2.5-72B were highly sensitive to UALD. Full-precision and 4-bit models showed broadly similar drift patterns, whereas 2-bit models behaved distinctly.
What the result does not establish
The result has a narrow scope. It comes from model-produced answers to a 60-item MBTI-style questionnaire under specified prompts and deterministic decoding rules, so it is a finding about behavior in this test rather than a general claim about personality.
The evaluation was single-turn rather than multi-turn. Its probability-based metrics do not provide causal interpretation, and the findings are specific to MBTI; whether they extend to other personality frameworks remains open.
Coverage was incomplete: the study did not include LLaMA3.1-405B or Qwen3-235B, and it did not test quantization-aware training or parameter-efficient fine-tuning. Whether the reported patterns hold at those scales or with those methods remains open.
Paper data and sources
Original title: When Personality Meets Quantization: A Layer-wise MBTI Analysis of Quantized LLMs
Authors: Yao Fu, Lijia Huang, Xiaomin Li et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text