Still, the researchers wanted to ensure that their model understood the instructions independently of the probe, instead of the probe inferring the robot's movements from the LLM’s grasp of syntax.

→ View original post on X — @mit_csail,