Researchers at an AI optimization lab called Fulcrum have found that language model agents are often 'under-elicited'—essentially leaving potential on the table. In a study shared on LessWrong, they tested simple interventions: telling models a task is solvable, suggesting better search strategies, and rotating fresh agents into a task when one gets stuck. The combination roughly doubled performance across multiple optimization challenges. The finding is striking for its simplicity: rather than building smarter models, significant gains may come from asking the right questions.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton