E-M-N-L-P

Abstract

Simulating Professions (SP) with Large Language Models (LLMs) has shown promising potential in education, healthcare, law, and other domains. However, the moral and ethical behavior of LLMs in educator roles remains underexplored. In this work, we introduce EMNLP (Educator-role Moral and Normative LLMs Profiling), a comprehensive framework for evaluating personality traits, moral development stages, and harmful content risk in teacher-role LLMs. EMNLP integrates an extended teacher personality scale (CPST-E), a diverse set of 88 moral dilemmas specific to the teaching context, and a soft prompt injection module targeting four moral flaw dimensions. Through experiments on 12 LLMs, we uncover key trends: teacher SP LLMs tend to exhibit idealized yet emotionally muted personality profiles, perform well in abstract ethical reasoning but poorly in emotionally nuanced scenarios, and show a tradeoff between reasoning strength and robustness against prompt-based role manipulation. Additionally, we evaluate the impact of temperature settings on behavioral consistency. Our results highlight the need for finer-grained alignment strategies to ensure ethical behavior in high-stakes educational simulations.

EMNLP: Educator-role Moral and Normative Large Language Models Profiling

Abstract

HEXACO Radar Chart comparing personality traits of teacher LLMs and real teachers at temperature=0. The visualization shows how model-generated teacher personalities differ from those of in-service teachers across six major personality dimensions.

HEXACO Radar Chart showing personality trait comparison at temperature=0.25. This visualization demonstrates how sampling temperature affects the personality profile generated by teacher LLMs compared to real teachers.

HEXACO Radar Chart at temperature=0.5, illustrating mid-range temperature effects on teacher LLM personality trait expression compared to real teacher averages across the six HEXACO dimensions.

HEXACO Radar Chart at temperature=0.75, showing how increased randomness in sampling affects the personality trait distribution of teacher LLMs compared to the more balanced profile of real-world educators.

HEXACO Radar Chart at maximum temperature=1.0, displaying the highest variance in personality trait expression among teacher LLMs. This visualization highlights how high-temperature sampling leads to more divergent personality profiles compared to real teachers.

Item-level comparison on the HEXACO-60 inventory at temperature=0. This detailed visualization breaks down how teacher LLMs respond to individual personality assessment items compared to real teachers, showing more polarized responses from the models.

Item-level comparison on the HEXACO-60 inventory at temperature=0.25, showing how slight randomness affects responses to individual personality assessment items across different LLMs compared to in-service teachers.

Item-level comparison on the HEXACO-60 inventory at temperature=0.5, illustrating medium-temperature effects on item-specific personality trait expressions across various teacher LLMs and real teacher averages.

Item-level comparison on the HEXACO-60 inventory at temperature=0.75, showing how higher randomness influences responses to specific personality items, with increasingly varied responses from teacher LLMs.

Item-level comparison on the HEXACO-60 inventory at maximum temperature=1.0, demonstrating how high randomness in sampling produces the most diverse responses to personality items from teacher LLMs compared to the more consistent responses of real teachers.

CPST-E Radar Chart at temperature=0, comparing professional teacher personality traits between LLMs and real teachers. This visualization focuses on teaching-specific personality dimensions that influence professional practice.

CPST-E Radar Chart at temperature=0.25, showing how low-temperature sampling affects the professional personality traits of teacher LLMs compared to in-service teachers across teaching-specific dimensions.

CPST-E Radar Chart at temperature=0.5, illustrating medium-temperature effects on professional teacher personality traits across various LLMs compared to real teacher averages.

CPST-E Radar Chart at temperature=0.75, showing how increased randomness affects the professional personality profile of teacher LLMs across teaching-specific dimensions compared to real educators.

CPST-E Radar Chart at maximum temperature=1.0, displaying the highest variance in professional teacher personality traits among LLMs. This shows how high-temperature sampling impacts teaching-specific attributes compared to real teachers.

Item-level comparison on the CPST-E inventory at temperature=0. This detailed visualization breaks down how teacher LLMs respond to individual professional personality assessment items compared to real teachers.

Item-level comparison on the CPST-E inventory at temperature=0.25, showing how slight randomness affects responses to professional teaching personality items across different LLMs compared to in-service teachers.

Item-level comparison on the CPST-E inventory at temperature=0.5, illustrating medium-temperature effects on teaching-specific personality traits across various LLMs and real teacher averages.

Item-level comparison on the CPST-E inventory at temperature=0.75, showing how higher randomness influences responses to specific professional teaching personality items from LLMs.

Item-level comparison on the CPST-E inventory at maximum temperature=1.0, demonstrating how high randomness in sampling produces the most diverse responses to professional teaching personality items from LLMs compared to real teachers.

BibTeX