Geometry of probe weights. For each 3D plot, one dot represents one tile on an Othello board, and neighbouring tiles have their two dots connected.
[Linear Probes trained on randomized Othello-GPT]
[Linear Probes trained on the Othello-GPT trained on championship dataset]
[Linear Probes trained on the Othello-GPT trained on synthetic dataset]
[Nonlinear Probes trained on randomized Othello-GPT]
[Nonlinear Probes trained on the Othello-GPT trained on championship dataset]
[Nonlinear Probes trained on the Othello-GPT trained on synthetic dataset]