Humanizing LLMs through Prompt Engineering

Imagine having a conversation with an AI that not only answers your questions but does so with the nuance, reasoning, and personality of a human. Thanks to breakthroughs in prompt engineering, this vision is becoming a reality. In our previous post, we explored the art of crafting effective prompts to unlock the power of LLMs. Now, we’ll take it a step further: how do we teach LLMs to think, talk, and even behave like humans?

‍

In this post, we’ll dive into three key techniques—Chain-of-Thought reasoning, model alignment, and personality engineering—which are transforming LLMs from mere text generators into intelligent, relatable conversational partners.

‍

Chain-of-Thought (CoT): Teaching LLMs how to Think like Humans

‍
If minor tweaks in phrasing can lead to wildly different outputs, how do we build reliable human-like cognition in LLMs using prompts? In their landmark paper “Large Language Models are Zero-Shot Reasoners,” Kojima et al demonstrated that appending “Let’s think step by step” to any prompt broadly increases the mathematical and reasoning abilities of language models. This prompt leads to CoT behavior since it encourages output with a sequence of short reasoning steps until the model can calculate a final answer. Their benchmarks show that this simple modification increases model accuracy across many common reasoning tasks, such as arithmetic and symbolic reasoning.

‍

Interestingly, this mirrors the two levels of human cognition. System 1 reasoning abilities are unconscious and driven by reflex. System 2 reasoning abilities are those requiring in-depth logical thought or calculation — those that require CoT. By appending “Let’s think step by step” to a prompt, we are able to unlock higher System 2 reasoning abilities in language models.

‍

Model Alignment: Teaching LLMs how to Talk like Humans

Whereas CoT reasoning teaches models to think like humans, model alignment teaches them to talk like us and hold natural conversations. It refers to the process of getting an LLM to provide answers in a desired format. For instance, while a base model like GPT-3.5 might respond with blunt, technical answers, its aligned conversational variant (ChatGPT) adds flourishes such as “Great question! Let me break this down…”

‍

The “LIMA: Less Is More for Alignment” paper showed that using only a thousand examples is enough to align an LLM to interact with users in a specific way. In “The Unlocking Spell on Base LLMs: Rethinking Alignment Via In-Context Learning,” researchers showed that only 3(!) examples allow us to align an LLM to interact with the user conversationally. They ran an interesting set of experiments where they compared a base LLM with its aligned counterpart, and showed that they usually agreed when asked to predict the next token. They only differed on a small set of stylistic tokens, such as hello/thank/please/additionally, which did not change the meaning of a sentence but significantly enhanced the user experience.

‍

The researchers crafted a set of three detailed examples showing a base LLM how to interact with the user conversationally, focused on the use of these special stylistic tokens. This approach leveraged In-Context Learning where LLMs exhibit an impressive ability to learn surprisingly well from just a few examples. When given these three detailed examples, an LLM then chats with the user in a natural conversational format.

‍

Personality Engineering: Teaching LLMs how to be like Humans

After enabling reasoning (CoT) and communication (alignment) in LLMs, we can use careful prompt engineering to install the final piece: personality. Researchers have demonstrated we can systematically program psychological traits into LLMs, creating distinct personalities. In “Personality Traits in Large Language Models,” researchers wrote detailed personality prompts for ChatGPT and then performed several standard tests to measure the Big Five personality traits: extraversion, agreeableness, conscientiousness, neuroticism and openness to experience. They revealed that different LLM personalities reliably have different Big Five personality traits.

‍

The researchers also proved that different prompts can reliably influence different aspects of LLM personality, so that for instance we can make an LLM more extraverted while keeping the rest of its personality intact. This shows that solely using prompt engineering can provide customizable user experiences, such as highly personalized tutors or therapists.

‍

Conclusion

‍
With the recent breakthroughs in Prompt Engineering, this field has evolved from simple input instructions to the primary tool used for transforming an LLM’s core cognitive processes. It shapes how the LLM thinks, how it talks, and ultimately, who it becomes. As research in this field advances, new prompt engineering techniques will further strengthen the most important partnership in tech — one between human and machine intelligence. 

‍