Does Progressive Training Improve Neural Network Reasoning Ability?

This section offers crucial empirical proof that Transformers cling to statistical short cuts rather than learning true reasoning algorithms from data. Graph connection tasks are used by the authors to illustrate this. Upon training on typical random graphs, a model rapidly reaches an accuracy of about 80%. Further research, however, shows that it does not learn pathfinding; instead, it just predicts "connected" by default and occasionally predicts "not connected" using a rudimentary heuristic that involves determining if the source or target node has a degree of zero.