Weight Direction, Not Magnitude, Drives Grokking in Neural Networks.

Researchers investigated the "grokking" phenomenon in neural networks, where models generalize long after overfitting training data. They developed a novel technique called "cross-trajectory chimera interventions," which involves combining the weight norms from one independently trained network with the weight directions from another. This allowed them to isolate the causal roles of these two weight properties. The study revealed that the direction of a network's weight vectors is crucial for determining which specific solution or "circuit identity" the network adopts. When a donor network's weight directions were implanted into a recipient network, the recipient consistently converged to the donor's solution. In contrast, the magnitude of the weights primarily affected how quickly a network learned and how susceptible its learned identity was to being overwritten, rather than dictating the identity itself.