Can Dialects Be Steered Like Languages? Sparse Neurons and Distributed Directions in Arabic LLMs

Can Dialects Be Steered Like Languages? Sparse Neurons and Distributed Directions in Arabic LLMs

Arabic language models exhibit dialect-specific neural representations that can be manipulated at inference time to control dialectal output without requiring additional training.

Hugging Face · Daily Papers ·Kareem Elozeiri, Mervat Abassy ·

Este artigo está em destaque na seleção diária de papers do Hugging Face, curada pela comunidade de pesquisa em IA.

Autores: Kareem Elozeiri, Mervat Abassy, Omar Kallas, Fahim Dalvi, Preslav Nakov, Kentaro Inui

  • 0 upvotes da comunidade
  • Temas: neuron-level analysis, sparse neuron populations, dialect-specific features, vector-steering approach, activation directions, dialect control

Resumo

Resumo original (em inglês), extraído do paper:

Arabic language models exhibit dialect-specific neural representations that can be manipulated at inference time to control dialectal output without requiring additional training.

Onde ler

compartilhar: