The Long Delay to Arithmetic Generalization: When Learned Representations Outrun Behavior

Researchers investigate why transformers experience prolonged training plateaus before sudden generalization ('grokking') on arithmetic tasks. Using causal interventions on Collatz prediction models, they show encoders learn structure early while decoders create a bottleneck preventing accuracy improvement. Numeral base choice dramatically affects convergence: bases aligned to task arithmetic reach 99.8% accuracy while binary fails completely.