Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Personality-Driven Decision-Making in LLM-Based Autonomous Agents

Lewis Newsham, Daniel Prince

TL;DR

This work addresses how prompt-induced OCEAN personality traits shape real-time decision-making in LLM-based autonomous agents for cyber defense, extending prior SANDMAN findings from planning to ongoing task selection. It presents a robust framework: 500 work schedules, a controlled persona induction schema (five traits × two directions plus baseline), and an analytical pipeline that treats schedules as sequences and uses movement deltas along with multiple sequence-similarity metrics to quantify trait-driven transformations. The study employs statistical tests with Bonferroni corrections to compare against baselines and investigates the influence of sampling temperature on determinism. Results show that inducible traits elicit plausible, trait-consistent shifts in task prioritisation, especially for Conscientiousness and Extraversion, with GPT-4o models displaying stronger effects than GPT-3.5-Turbo, highlighting both the potential for controllable, believable Deceptive Agents and the need for careful ethical safeguards in such systems. Overall, the paper provides a quantitative, generalizable method to evaluate persona-driven decision-making in LLM-based agents and underscores considerations for practical deployment in cyber defense contexts.

Abstract

The embedding of Large Language Models (LLMs) into autonomous agents is a rapidly developing field which enables dynamic, configurable behaviours without the need for extensive domain-specific training. In our previous work, we introduced SANDMAN, a Deceptive Agent architecture leveraging the Five-Factor OCEAN personality model, demonstrating that personality induction significantly influences agent task planning. Building on these findings, this study presents a novel method for measuring and evaluating how induced personality traits affect task selection processes - specifically planning, scheduling, and decision-making - in LLM-based agents. Our results reveal distinct task-selection patterns aligned with induced OCEAN attributes, underscoring the feasibility of designing highly plausible Deceptive Agents for proactive cyber defense strategies.

Personality-Driven Decision-Making in LLM-Based Autonomous Agents

TL;DR

This work addresses how prompt-induced OCEAN personality traits shape real-time decision-making in LLM-based autonomous agents for cyber defense, extending prior SANDMAN findings from planning to ongoing task selection. It presents a robust framework: 500 work schedules, a controlled persona induction schema (five traits × two directions plus baseline), and an analytical pipeline that treats schedules as sequences and uses movement deltas along with multiple sequence-similarity metrics to quantify trait-driven transformations. The study employs statistical tests with Bonferroni corrections to compare against baselines and investigates the influence of sampling temperature on determinism. Results show that inducible traits elicit plausible, trait-consistent shifts in task prioritisation, especially for Conscientiousness and Extraversion, with GPT-4o models displaying stronger effects than GPT-3.5-Turbo, highlighting both the potential for controllable, believable Deceptive Agents and the need for careful ethical safeguards in such systems. Overall, the paper provides a quantitative, generalizable method to evaluate persona-driven decision-making in LLM-based agents and underscores considerations for practical deployment in cyber defense contexts.

Abstract

The embedding of Large Language Models (LLMs) into autonomous agents is a rapidly developing field which enables dynamic, configurable behaviours without the need for extensive domain-specific training. In our previous work, we introduced SANDMAN, a Deceptive Agent architecture leveraging the Five-Factor OCEAN personality model, demonstrating that personality induction significantly influences agent task planning. Building on these findings, this study presents a novel method for measuring and evaluating how induced personality traits affect task selection processes - specifically planning, scheduling, and decision-making - in LLM-based agents. Our results reveal distinct task-selection patterns aligned with induced OCEAN attributes, underscoring the feasibility of designing highly plausible Deceptive Agents for proactive cyber defense strategies.

Paper Structure

This paper contains 18 sections, 1 equation, 8 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Background and Related Work
  3. Method
  4. Experimental Process
  5. Prompt-Based Persona Induction
  6. Analytical Approach
  7. Results
  8. Plausibility Analysis
  9. Conscientiousness: Work-related Activities
  10. Movement Delta Shift: Work
  11. Extraversion: Social Activities
  12. Transformation Analysis
  13. Effect of Sampling Temperature on Sequence Alignment
  14. Transformation Significance: Experimental Conditions vs. Control
  15. GPT-4o and GPT-4o-Mini
  16. ...and 3 more sections

Figures (8)

  • Figure 1: Decision-making task to be performed by the LLM featuring Agreeableness (Positive) as the induced trait.
  • Figure 2: Movement Deltas: Positive Conscientiousness (GPT-4o).
  • Figure 3: Movement Deltas: Negative Conscientiousness (GPT-4o).
  • Figure 4: Movement Deltas: Work (GPT-4o & GPT-3.5-Turbo)
  • Figure 5: Movement Deltas: Positive Extraversion (GPT-4o).
  • ...and 3 more figures