The Power of Questions: How GATE enhances AI’s ability to read the human mind

We’ve all been there, haven’t we? Whether you’ve worked in a customer-facing job or collaborated with a diverse team, you’ve likely encountered the puzzle of everyone’s unique preferences. Understanding what each individual wants can be challenging, even for us humans. But what about Artificial Intelligence (AI) models? They lack the direct human experience we have to draw upon and understand what people are looking for.

Enter a group of researchers from renowned institutes and Anthropic, the creators of the large language model (LLM) known as Claude 2. They’re tackling this challenge with a seemingly straightforward yet ingenious solution: encouraging AI models to engage users with questions to uncover their true desires.

Welcome to a new dimension of AI understanding through GATE

In a recent research paper, startup Anthropic’s Alex Tamkin, along with colleagues Belinda Z Li and Jacob Andreas from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL), and Noah Goodman from Stanford, introduced Generative Active Task Elicitation (GATE). Their goal? To empower language models to translate human preferences into automated decision-making systems.

In other words, they harness the LLM’s ability to analyse and generate text to initiate a dialogue with the user during their first interaction. The LLM then adapts its responses in real-time based on the user’s input and, importantly, deduces the user’s underlying needs by analyzing its extensive database.

The three facets of GATE

Generative active learning: Here, the LLM showcases the kind of responses it can provide and seeks the user’s feedback. For instance, it might ask, “Are you interested in the following article: ‘The Art of Fusion Cuisine: Mixing Cultures and Flavors’?” Depending on the user’s response, the LLM fine-tunes its subsequent content.

Yes/no question generation: This method involves the LLM posing binary yes or no questions, like, “Do you enjoy reading articles about health and wellness?” It then tailors its future responses based on the user’s answers, steering clear of topics associated with a “no” response.

Open-ended questions: Similar to the first approach but broader in scope, the LLM aims to extract the most abstract knowledge from the user. It might ask questions like, “What hobbies or activities do you enjoy in your free time, and why do these interests fascinate you?”

Promising outcomes

The researchers put the GATE method to the test in three domains: content recommendation, moral reasoning, and email validation. By refining OpenAI’s GPT-4 with GATE and involving 388 paid participants to answer questions and evaluate responses, the researchers found that GATE often led to more accurate models compared to traditional methods, all while requiring similar or even less cognitive effort from users.

In particular, they noted that GPT-4 fine-tuned with GATE exhibited a significant improvement of about 0.05 points in subjective measurements when it came to understanding individual preferences. This might sound small, but it’s a substantial leap, especially when starting from scratch.

The researchers assert that they’ve presented compelling initial evidence that language models can effectively employ GATE to comprehend human preferences with greater accuracy and less user effort than conventional methods.

This innovation has the potential to save considerable time for enterprise software developers when implementing LLM-powered chatbots for customer or employee-facing applications. Instead of relying on pre-existing data to understand individual preferences, fine-tuning the models with the GATE method could lead to more engaging and helpful experiences for users.

So, if your go-to AI chatbot starts quizzing you about your preferences in the future, chances are it’s using the GATE method to provide you with even better responses.