In-Context Impersonation Reveals Large Language Models' Strengths and Biases (2305.14930v2)

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.

In-Context Impersonation in LLMs: A Detailed Study on Strengths and Biases

The paper under review, titled "In-Context Impersonation Reveals LLMs' Strengths and Biases," provides a thorough investigation into how LLMs can adapt their performance across various tasks by impersonating different personas. The authors, Salewski et al., have designed experimental scenarios to probe both the capabilities and the limitations of LLMs when tasked with role-based prompts. Here we provide a structured analysis of their results, methodologies, and implications.

Methodology

The primary methodology employed involves prefixing LLM prompts with specified personas. The personas encapsulate distinct social identities or domains of expertise, thereby directing the LLM to perform tasks as if it were a person from that domain or identity. The paper evaluates this approach using three separate tasks: a multi-armed bandit task, a reasoning task using the MMLU dataset, and a vision-language task incorporating fine-grained image classification.

1. Multi-Armed Bandit Task: LLMs were prompted to impersonate children of varying ages, to analyze how exploration strategies evolve akin to human developmental stages. This simulated the psychological tendency for younger beings to explore more than older ones.
2. Reasoning Task: This involved presenting the LLMs with questions requiring reasoning from the MMLU dataset. The impersonations varied among domain-specific personas, assessing improvements in task performance when LLMs are cued to adopt expert roles.
3. Vision-Language Task: Here, the focus was on using persona-based textual descriptions to enhance fine-grained visual classification tasks. Performance improvements were noted when domain-relevant expert personas were employed, but biases were also evident when examining social identities.

Results and Observations

The findings from these experiments are significant, revealing both the adaptive strengths and underlying biases of LLMs:

• Multi-Armed Bandit Task Results: LLMs showed differential performance based on impersonated age. Models impersonating older ages garnered higher average rewards, reflecting a transition from exploration-focused to exploitation-focused strategies—a pattern consistent with human developmental psychology.
• Reasoning Task Results: The impersonation of domain experts significantly enhanced model performance compared to non-expert personas. This suggests that LLMs, when guided by contextual cues of expertise, can align their outputs more closely with expert-level reasoning.
• Vision-Language Task Results: Impersonating expert roles improved classification tasks, but significant biases were observed—such as gender and racial biases being reflected in the outputs. These biases highlight the models' inclination to embody stereotypes, thereby raising ethical concerns regarding their application in sensitive contexts.

Implications and Future Directions

The investigation offers several implications for the future of AI and LLMs:

• Practical Implications: The findings encourage the use of role-based prompts to enhance LLM task performance in domains requiring specialist knowledge. However, this also necessitates a closer examination of ethical guidelines to mitigate biased outputs when impersonating social identities.
• Theoretical Implications: The ability of LLMs to exhibit human-like learning patterns and biases when prompted in context opens new avenues in cognitive modeling using artificial systems, providing insights into the design of models that might simulate human-like reasoning.
• Future Developments in AI: Further research could explore scaling the approach to more complex, interactive tasks and understand how varied training data and fine-tuning impact the observed biases. Investigating multilayered persona composition might also reveal nuanced insights about interrelations between different social or expert identities.

In conclusion, while in-context impersonation accentuates some strengths of LLMs, it also exacerbates some inherent biases, pointing to a dual-edged potential in practical applications. This paper serves as a crucial contribution to understanding the complexities involved when LLMs emulate human characteristics in their learning and output generation processes.