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

2026-02-16 · Source: cs.CL updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Advanced, extended

Summary

A systematic case study reveals that demographic-based persona assignments can significantly degrade the performance of Large Language Model (LLM) agents across diverse operational domains. Evaluating models like GPT-4o-mini, DeepSeek-V3, and Qwen3-235B on benchmarks spanning strategic reasoning, planning, and technical operations, researchers found performance degradations of up to 26.2%. These shifts are driven by task-irrelevant persona cues, indicating that simple prompt injections can distort an agent's decision-making reliability. The study assigned 23 personas across gender, race/origin, religion, and profession, observing that high-level reasoning tasks were more susceptible to disruption than technical tasks. For instance, DeepSeek V3's Card Game accuracy decreased by 26.2% under an "from Africa" persona, and gender roles influenced household planning tasks in ways reflecting societal stereotypes.

Key takeaway

For AI scientists and ML engineers deploying LLM agents in high-stakes environments, you must rigorously test your systems for persona-induced biases. Your agent's reliability can be compromised by seemingly innocuous demographic role assignments, leading to significant performance degradation. Implement robust validation processes to ensure task execution remains stable and equitable, independent of user-specified personas, especially for strategic reasoning and planning tasks where vulnerabilities are highest.

Key insights

Demographic persona assignments introduce implicit biases, significantly degrading LLM agent performance on action-oriented tasks.

Principles

• Persona conditioning affects LLM agent behavior beyond text generation.
• High-level reasoning tasks are more vulnerable to persona-induced bias.
• Implicit biases can mirror human social stereotypes.

Method

LLM agents were evaluated on five benchmarks (ALFWorld, WebShop, Card Game, OS Interaction, Database) using a fixed two-turn conversational prefix to assign 23 demographic-based personas, comparing performance against a no-persona baseline.

In practice

• Avoid demographic persona assignments in LLM agent prompts.
• Test LLM agents for bias across diverse demographic personas.
• Prioritize robustness in high-stakes agentic deployments.

Topics

• LLM Agents
• Persona Bias
• Agent Robustness
• Performance Degradation
• Strategic Reasoning

Best for: Machine Learning Engineer, NLP Engineer, AI Scientist, AI Researcher, AI Engineer, MLOps Engineer

Related on AIssential

• Study: AI chatbots provide less-accurate information to vulnerable users — AI chatbots underperform for vulnerable users, providing less accurate information and exhibiting biased behaviors.
• LLM-enabled Social Agents — Persona-based role definitions are crucial for LLM-enabled social agents to achieve socially intelligible and coherent behavior.
• The Self-Correction Illusion: LLMs Correct Others but Not Themselves — LLMs' self-correction failure is a chat-template artifact, not a cognitive deficit, showing role-label dependence.
• AgentNoiseBench: Benchmarking Robustness of Tool-Using LLM Agents Under Noisy Condition — LLM agents universally suffer significant performance degradation under realistic noise, especially from tool failures.
• RoleCDE:Benchmarking and Mitigating Role-Alignment Trade-offs in Role-Playing Agents — RoleCDE reveals LLM role-playing agents decouple from role-specific values when conflicting with alignment, a gap mitigated by targeted fine-tuning.
• Network Effects and Agreement Drift in LLM Debates — LLM agent opinion dynamics are shaped by network structure, group imbalance, and a systematic "agreement drift" bias.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CL updates on arXiv.org.