Precision Shapes Personality: The Hidden Cost of Quantization in Sub-Billion-LLMs

Soham Sen, Ishaan Gangwani

Keywords: Quantization, Post-Training Quantization (PTQ), Precision Effects, LLM Personality, Psychometric Evaluation, TRAIT Benchmark, BIG5 (OCEAN), Short Dark Triad (SD3), Extraversion, Conscientiousness, Narcissism, Openness, Machiavellianism, Numeric Stability, Sub-Billion LLMs, On-Device Deployment, NF4 Quantization, INT8 Quantization, Edge AI, Deterministic Evaluation, Empathy and Safety in LLMs

Abstract: Psychometric studies of language models are increasingly important given their growing use as human assistants and in therapeutic settings. Such applications are often deployed on edge devices with sub-billion parameter large language models (LLMs) operating under strict memory and latency constraints, where post-training quantization (PTQ) is standard. Yet little is known about whether numeric precision alters measured personality traits. In the current work, using a psychometric benchmark TRAIT, we evaluate five sub-1B LLMs across different precision settings. We find that 4-bit Normal Float (nf4) produces the largest shifts, int8 smaller ones, and 16-bit formats remain closest to native. Shifts concentrate in Extraversion, Conscientiousness, and Narcissism, while Openness and Machiavellianism are more stable. These results identify precision as a consequential, controllable variable that should be disclosed and audited when personality matters in deployment.

Submission Number: 241