Building Better AI Agents: A Provocation on the Utilisation of Persona in LLM-based Conversational Agents

📝 Paper Summary

Agentic simulation Conversational personalization Anthropomorphism

This position paper argues that LLM-based conversational agents must transition from generic personality traits to distinct, consistent 'personas' to effectively serve in simulation, professional roles, and branding.

Core Problem

Current LLM-based agents often lack a distinct 'soul' or consistent identity, relying instead on generic personality traits (e.g., 'friendly') or prompt-based personas that degrade into inconsistency or hallucination over long conversations.

Why it matters:

Inconsistent personas break user trust and engagement, particularly in long-term interactions like healthcare or education
Generic 'personality' traits fail to capture the nuance required for specific character simulations (e.g., simulating a specific demographic for social science research)
Malicious actors can exploit persona inconsistencies to generate toxic responses or break character constraints

Concrete Example: A CA assigned generic traits like 'bravery' and 'smartness' might alternate between acting like Sherlock Holmes and Hermione Granger depending on the context. Similarly, a '21-year-old student' persona might claim to like bowling in one turn but prefer skiing in the next, revealing a lack of consistent identity.

Key Novelty

Distinction between Persona and Personality in LLM Agents

Formalizes the distinction where 'Personality' refers to generic traits (friendly, smart) shared by groups, while 'Persona' represents a complex, consistent identity (history, tone, voice) of a unique character
Identifies 'Persona Consistency' and 'Persona Hallucination' (holding wrong beliefs about one's own identity) as the primary technical hurdles preventing realistic agent deployment

Breakthrough Assessment

4/10

This is a provocation/survey paper rather than a technical breakthrough. It usefully frames the problem of 'persona' distinct from 'personality' but does not propose a new architecture or solve the consistency issues it identifies.

⚙️ Technical Details

Problem Definition

Setting: Design and evaluation of Conversational Agents (CAs) that maintain a specific, complex identity over multi-turn dialogues

Inputs: Natural language prompts defining a character (e.g., demographics, history, tone)

Outputs: Conversational responses that adhere to the defined persona constraints without drifting or hallucinating attributes

Comparison to Prior Work

vs. General LLMs: General models prioritize broad applicability and helpfulness, often lacking the specific constraints required for 'soul' or identity
vs. Personality Embeddings (Li et al., 2016): Prior work embedded personas as simple vectors; this paper argues for a more comprehensive integration involving domain knowledge and consistency checks

Limitations

Lack of systematic evaluation frameworks: Current methods rely on psychometric tests or surveys which cannot fully capture persona consistency
Inconsistency vulnerability: Most open-source LLMs struggle to align responses with latent persona attributes, especially under complex questioning or negation
Persona Hallucination: Agents may confidently adopt incorrect facts about their assigned persona that are difficult to correct via standard hallucination detection methods
Ethical risks: Anthropomorphism can lead to deception, reinforcement of stereotypes, and manipulation of user emotions

Reproducibility

No replication artifacts mentioned in the paper. This is a position paper citing existing literature; no new models or datasets were introduced.

📊 Experiments & Results

Main Takeaways

Current LLMs (like ChatGPT) are not designed with distinct personas, leading to 'mechanical' interactions; assigning personas via prompts is possible but technically fragile
Consistency is the top priority challenge: agents easily drift from their assigned persona (e.g., changing age or hobbies) during extended or complex conversations
Persona is critical for specific applications: Participant Simulation (replicating social science experiments), Role Playing (training doctors/lawyers), and Brand Representation require consistent identities, not just generic traits
Domain knowledge is inseparable from persona: A 'judge' persona must know the law, not just sound authoritative; current approaches often neglect this knowledge integration

📚 Prerequisite Knowledge

Prerequisites

Understanding of LLM prompting strategies (system prompts)
Familiarity with hallucinations in LLMs
Basic concepts of anthropomorphism in HCI (Human-Computer Interaction)

Key Terms

CA: Conversational Agent—a dialogue system designed to communicate with humans in natural language

Persona: The essence or 'soul' of an agent, encapsulating distinct tone, voice, history, and static characteristics (race, age) to form a unique identity, distinguished here from generic personality traits

Personality: Generic traits (e.g., 'friendly', 'smart') or frameworks like MBTI that describe characteristics shared by groups, rather than a specific unique identity

Persona Consistency: The ability of an agent to maintain unchanged persona elements (e.g., age, history) throughout a multi-turn conversation without contradiction

Persona Hallucination: A failure mode where an agent with a defined persona holds wrong beliefs about itself (e.g., wrong occupation or social relationships) or generates facts inconsistent with its assigned character

LLM: Large Language Model—AI systems trained on vast amounts of text to generate human-like language