The team first developed a set of small Karel puzzles, which involve coming up w/instructions to control a robot in a simulated environment. Then, they trained an LLM on the solutions w/o showing it how those solutions actually worked.

→ View original post on X — @mit_csail,