One paper accepted to AAAI 2025

Usually, they run experiments where they give one group the new treatment and another group nothing. But these experiments are expensive and time-consuming — so researchers are always looking for better ways to run them.

This is where ABC3, a smart algorithm developed by researchers at Korea University, comes in.

What’s the Problem?

Traditional experiments treat all participants equally — like flipping a coin to decide who gets what. But in real life, some participants provide more useful information than others. Picking them wisely could save a lot of time and money.

The challenge is to choose participants in a way that doesn’t mess up the scientific fairness of the experiment (called “randomization”) and still gives reliable results.

Meet ABC3: An AI-Powered Assistant for Experiments

ABC3 is like a thoughtful assistant for scientists. It uses Bayesian statistics — a branch of math that helps update beliefs based on new data — to pick the best next person to include in the study.

It does this by:

• Reducing uncertainty about how the treatment affects different people,
• Keeping the experiment fair by ensuring both treated and untreated groups are balanced,
• And avoiding false positives, where a treatment looks effective by accident.

All of this is done using a clever method inspired by a 1990s approach called the “Cohn Criterion,” which helps machines learn from data more efficiently.