Frontier Reasoning Agents Fail on Interactive 2D Mazes (multinet.ai)

There's a lot of hype out there but it seems like even frontier models still aren't that good at control tasks... And this is not a sim to real issue because these aren't real world tests. We likely need a different model architecture to address control

Hey HN, I’m one of the co-authors of this work. Very excited to see it here

We’ll actually be publishing a full paper soon, but happy to answer any questions in the meantime!