Thanks, still finishing the blog post so I'll cover all that! It's designed closer to a continual learning benchmark, new context for each problem and models get to transfer their lessons to future selves. They also have access to prior solutions, as that handoff is necessary to

→ View original post on X — @alexjc,