Understanding Intent Through Conversation with LLMs

Why prompts can be difficult

Even when people are asked to write down what they want in a prompt, it can be difficult to describe their preferences clearly in language. Users may not fully know what they want when assigning a task to an LLM, and may also struggle to state it precisely.

This research suggests that, instead of requiring a complex prompt, an LLM can clarify a user’s intent by asking questions and exchanging examples. The central idea is that the model can elicit a person’s intended preferences through conversation.

GATE: Generative Active Task Elicitation

The paper, “Eliciting Human Preferences with Language Models,” by Belinda Z. Li, Alex Tamkin, Noah Goodman, and Jacob Andreas, proposes the GATE approach: Generative Active Task Elicitation.

GATE is positioned between two existing approaches. Conventional prompting is free-form but passive: the user provides an instruction and the model responds. Pool-based active learning is example-based and active: the system selects examples to ask about. GATE is both free-form and active, using generated questions and examples to identify the user’s preferences.

• Conventional prompting: free-form and passive
• Pool-based active learning: example-based and active
• GATE: free-form and active

What the results suggest

Figure 3 indicates that GATE can identify human preferences effectively. Figure 4 indicates that it imposes the same or a lower cognitive burden on users.

If this direction develops further, prompt engineering—currently emphasized as an important skill—may eventually become far less necessary. It may be like driving: someone can become an excellent driver without knowing how to operate a manual transmission.

Original paper

“Eliciting Human Preferences with Language Models” (2023). Original paper: https://arxiv.org/abs/2310.11589