A video test appears to show a model backtracking through a stuck-puck problem

A post points to a reliability test later in a video as an example of what it calls “physical reasoning”: when something doesn’t fit, the model backtracks and tries a different approach. In the puck example, it removes a stuck puck, makes space elsewhere and puts it back. The poster argues that step-by-step problem-solving like this should work with models out of the box.