Selective Neuron Amplification in Transformer Language Models

Ryyan Akhtar, Payal Pahwa, Monika Arora

Large language models often fail on tasks they seem to already understand. In our experiments, this appears to be less about missing knowledge and more about certain internal circu…