The Long Delay to Arithmetic Generalization: When Learned Representations Outrun Behavior
A new research paper explores the phenomenon of 'grokking' in transformers, where models abruptly generalize after a long delay during training on algorithmic tasks. The study suggests this delay stems from limited access to learned structures rather than an inability to acquire them. By analyzing one-step Collatz prediction, researchers found that while encoders quickly learn relevant structures, the decoder bottleneck prolongs the generalization phase. Interventions like transplanting trained encoders or freezing encoders and retraining decoders significantly accelerated learning and improved accuracy, with numeral representation also playing a crucial role. AI
IMPACT Provides insights into transformer learning dynamics, potentially informing future model architectures and training strategies for improved efficiency.