More Options Don't Make AI Smarter — New Study Explains Why

Researchers studied a popular AI training method called Option-Critic. In this method, an AI isn't just taught one way to act. It's given multiple 'options' — like little playbooks for different situations — and rules for when to switch between them. The method had a headline result: performance improves as you add more options. But this new paper explains that this improvement isn't because the options themselves are clever.

First, they found that the rule AI learns for when to switch options is basically useless. In practice, it can actually block AI from exploring new choices, making it much slower to learn. The authors show that simply forcing AI to switch constantly works just as well — completely removing the need for a learned switching rule.

Second, they identified a problem they call 'policy necrosis' — like tissue death. An AI state locks onto the first action that looked promising and then never updates. In their experiments, three out of five states were 'necrotic.' This means AI ignores better solutions. When they restored exploration, just one option could solve a whole task.

Finally, why do extra options help at all? The study shows it's not because any individual option gets better. It's because having more options reduces the chance that all of them fail at the same time — from 59% down to 4%. So adding options is like buying more lottery tickets, not making smarter guesses. The real fix is better exploration.