From Biased Chatbots to Biased Agents: Examining Role Assignment Effects on LLM Agent Robustness

1. Problem Statement

While Large Language Models (LLMs) are increasingly deployed as autonomous agents capable of executing real-world actions (e.g., code deployment, medical decisions, financial trading), research has largely focused on biases in text generation rather than action execution.

• The Gap: It is unknown whether demographic-based persona assignments (e.g., assigning a specific gender, race, or profession) affect an agent's ability to perform multi-step operational tasks.
• The Risk: If persona cues introduce implicit biases that alter decision-making reliability, it poses a direct operational risk in high-stakes environments where agents must remain robust regardless of user-specified context.

2. Methodology

The authors conducted the first systematic case study investigating how demographic personas impact LLM agent performance across diverse operational domains.

Experimental Setup

• Models Evaluated: Three widely deployed LLMs representing different architectures and origins:
  • GPT-4o-mini: Commercial model (OpenAI).
  • DeepSeek-V3: Open-source model.
  • Qwen3-235B: Open-source Mixture-of-Experts (MoE) model.
• Persona Categories: 23 distinct personas spanning four demographic dimensions:
  • Gender: Female, Male, Non-Binary.
  • Race/Origin: Black, White, Asian, from Africa, from Europe, from America.
  • Religion: Christian, Muslim, Hindu, Buddhist, Jewish, Chinese Traditional.
  • Profession: Professor, Doctor, Manager, Student, Farmer, Laborer, Developer, CEO.
• Prompting Strategy: A fixed two-turn conversational prefix was used to condition the model into a specific persona before the task instruction. The baseline condition omitted this prefix entirely.
• Evaluation Benchmarks (5 Domains):
  1. ALFWorld: Embodied household task planning (Success rate).
  2. WebShop: E-commerce decision-making (Reward scores).
  3. Card Game: Strategic reasoning and planning (Win rate/Score).
  4. OS Interaction: System-level command execution (Command accuracy).
  5. Database: SQL query generation (Query correctness).

3. Key Contributions

• First Systematic Investigation: This is the first study to demonstrate that demographic persona assignments measurably degrade LLM agent performance in action-taking contexts, moving beyond text-only bias analysis.
• Identification of a New Vulnerability: The paper reveals that persona conditioning acts as a simple prompt injection vector that can distort an agent's decision-making reliability, introducing implicit biases that mirror human social stereotypes.
• Cross-Domain Evidence: The study proves that these effects are not limited to a single task type but span strategic reasoning, planning, and technical operations.

4. Key Results

The study found that persona assignments caused performance variations ranging from minor fluctuations to severe degradation (up to 26.2%), even though the personas were irrelevant to the tasks.

Quantitative Findings

• Maximum Degradation: DeepSeek V3's performance on the Card Game (strategic reasoning) dropped by 26.2% when assigned the "from Africa" persona.
• Model Variability:
  • GPT-4o-mini: Showed up to a 19.1% drop in Card Game performance with the "from America" persona.
  • Qwen3-235B: Experienced a 23.8% drop in Card Game accuracy when assigned the "White" persona.
• Task Sensitivity:
  • High Vulnerability: Tasks requiring multi-step reasoning and planning (Card Game, ALFWorld) were most susceptible to persona-induced disruption.
  • Relative Stability: Technical tasks (OS Interaction, Database) remained more stable, typically fluctuating within 2–5%.

Qualitative Patterns (Stereotype Correlation)

• Race/Origin: Assigning racial or geographic origins led to significant performance distortions. For example, "Asian" and "from Africa" personas severely degraded strategic reasoning in DeepSeek V3.
• Gender: Effects were task-dependent and reflected societal stereotypes.
  • Male personas degraded performance in household planning (ALFWorld), aligning with stereotypes about domestic competence.
  • Female and Non-Binary personas improved ALFWorld performance but sometimes degraded technical tasks.
• Profession: Performance correlated with perceived social status.
  • High-status roles (e.g., Doctor, CEO) often improved performance in some models.
  • Working-class roles (e.g., Laborer) consistently reduced performance in GPT-4o-mini.
• Religion: Religious identities caused volatile shifts. Christian and Buddhist personas significantly lowered Card Game accuracy in DeepSeek V3, while Jewish and Chinese Traditional roles improved it. This suggests model-specific pretraining data influences how religious cues are interpreted.

5. Significance and Implications

• Operational Risk: The findings expose a critical vulnerability in current agentic systems: implicit biases introduced via persona conditioning can undermine reliability. This is particularly dangerous in high-stakes fields like healthcare, finance, and justice, where performance must be consistent regardless of user context.
• Mechanism of Failure: The study suggests that LLM agents internalize socio-cognitive associations that interfere with core reasoning processes. The bias is not merely a change in tone but a fundamental alteration in the agent's ability to execute complex, multi-step plans.
• Future Directions: The paper calls for:
  1. Expanding evaluations to richer, collaborative multi-agent environments.
  2. Developing interpretability techniques to understand how persona conditioning alters action policies.
  3. Designing targeted debiasing interventions to ensure agents remain robust against demographic prompt injections.

Conclusion: As LLM agents transition from chatbots to operational executors, the "persona" is no longer just a stylistic choice but a potential vector for security and reliability failures. Addressing these persona-induced biases is essential for the safe deployment of autonomous AI systems.