When should an AI change its mind?

We study how AI can learn from sound advice without being led astray by recommendations that are wrong.

Why this question matters

Changing an answer can correct a mistake—or introduce one. We are interested in the distinction: what helps a model assess the substance of advice, rather than simply follow it?

What we are exploring

Controlled tasks let us compare responses to sound and unsound advice. One line of work examines whether checking task facts before seeing a recommendation changes the final decision.

Experimental research

Where the work stands

This is experimental work on observable behavior. Preliminary observations do not establish a general improvement or a mechanism inside the model.

Research · projects · collaboration

What are you thinking about?

Bring a question, a practical problem, or an idea you would like to explore together.

[email protected]