Path Channels and Plan Extension Kernels: a Mechanistic Description of Planning in a Sokoban RNN

ArXi:2506.10138v3 Announce Type: replace-cross We partially reverse-engineer a convolutional recurrent neural network (RNN) trained with model-free reinforcement learning to play the box-pushing game Sokoban. We find that the RNN s future moves (plans) as activations in particular channels of the hidden state, which we call path channels. A high activation in a particular location means that, when a box is in that location, it will get pushed in the channel's assigned direction.