MLST

Polysemanticity in neural networks arises from superposition, where a single neuron activates for multiple distinct inputs due to insufficient neurons. In language models, this enables efficient repre…