Sitemap

Member-only story

Divide and Conquer: How AI Learned to Ask the Right Questions — One Half at a Time”

Reimagining How Machines Talk About What They See

5 min readJul 31, 2025

--

Press enter or click to view image in full size
How can we teach AI to ask the right questions? Image created with DALL-E.

Imagine standing in front of a crowded shelf at the supermarket and trying to explain to someone over the phone how to find the exact product you want. You describe its color, shape, perhaps even its position. They ask, “Is it a box of cereal?” “Is it red?” “Is it on the top shelf?” And slowly, through back-and-forth questions, they narrow it down — until finally, they say, “Is it the red box of granola?” and you respond, “Yes!” Congratulations: you’ve just engaged in what AI researchers call goal-oriented visual dialogue.

This intuitive human strategy of visually guided Q&A has captivated the attention of AI scientists for years, particularly because of its implications in robotics, autonomous systems, and human-computer interaction. Yet despite numerous advances, one major challenge has remained unsolved: How can AI learn to ask smart, targeted questions that reduce uncertainty — fast?

A recent breakthrough, presented at AAAI 2025 by a team of Chinese researchers, offers a compelling answer. Their system, aptly named TSADE (Tree-Structured Strategy with Answer Distribution Estimator), mimics the way humans split complex problems into manageable halves. It’s a “divide…

--

--

Andreas Maier
Andreas Maier

Written by Andreas Maier

I do research in Machine Learning and head a Research Lab at Erlangen University, Germany.