[Paper Review] In-Context Impersonation Reveals Large Language Models' Strengths and Biases

Leonard Salewski, Stephan Alaniz+3 more|arXiv (Cornell University)|May 24, 2023

Language and cultural evolutionSocial Sciences35 citations

TL;DR

The paper shows that prompting LLMs to impersonate different ages, domains, and demographics in-context can both boost task performance and reveal biases across bandit, reasoning, and vision-language tasks.

ABSTRACT

In everyday conversations, humans can take on different roles and adapt their vocabulary to their chosen roles. We explore whether LLMs can take on, that is impersonate, different roles when they generate text in-context. We ask LLMs to assume different personas before solving vision and language tasks. We do this by prefixing the prompt with a persona that is associated either with a social identity or domain expertise. In a multi-armed bandit task, we find that LLMs pretending to be children of different ages recover human-like developmental stages of exploration. In a language-based reasoning task, we find that LLMs impersonating domain experts perform better than LLMs impersonating non-domain experts. Finally, we test whether LLMs' impersonations are complementary to visual information when describing different categories. We find that impersonation can improve performance: an LLM prompted to be a bird expert describes birds better than one prompted to be a car expert. However, impersonation can also uncover LLMs' biases: an LLM prompted to be a man describes cars better than one prompted to be a woman. These findings demonstrate that LLMs are capable of taking on diverse roles and that this in-context impersonation can be used to uncover their hidden strengths and biases.

Motivation & Objective

• Investigate whether LLMs can impersonate different social identities or domain experts in-context.
• Examine how impersonation affects performance on a two-armed bandit task.
• Evaluate reasoning performance under persona-based impersonation using the MMLU dataset.
• Assess how impersonation-generated descriptions influence downstream vision-language classification.
• Identify biases revealed by impersonation across gender and race in different tasks.

Proposed method

• Prompt LLMs with a prefix like "If you were a {persona}" to induce impersonation.
• Use a zero-shot setting with Vicuna-13B and, for reasoning, ChatGPT (gpt-3.5-turbo) across tasks.
• Bandit task: two-armed bandit with Bayesian updating and probit regression to analyze exploration (β1, β2).
• Reasoning task: multiple-choice questions from MMLU with task, domain, and non-domain expert personas.
• Vision-language task: generate class descriptions from personas and classify with CLIP using generated descriptions rather than class names.
• Evaluate across age-based (2,4,7,13,20) and domain/gender/race personas, repeating experiments for stochasticity.

Experimental results

Research questions

• RQ1 Can LLMs reproduce human-like exploration strategies by impersonating different ages in a bandit task?
• RQ2 Do domain-expert impersonations improve reasoning performance on MMLU compared to non-domain or neutral personas?
• RQ3 Do impersonation-generated descriptions enhance fine-grained visual classification via CLIP, and what biases appear across gender or race?
• RQ4 To what extent do age, expertise, ethnicity, and gender in prompts reveal biases in LLMs across tasks?

Key findings

• Older impersonation ages yield higher rewards and more exploitation in bandit tasks, mirroring human developmental patterns.
• Task-expert impersonation improves MMLU accuracy more than domain or non-domain impersonations.
• Impersonation-generated descriptions improve CLIP-based fine-grained classification for birds and cars, with performance increasing with age.
• Gender and race prompts reveal biases: e.g., men describe cars better and women describe birds better, with race effects also evident.
• Results are consistent across two LLMs (Vicuna-13B and ChatGPT) and several vision-language backbones, indicating robust effects of impersonation.