New method uses Optimal Transport to efficiently scale LLM depth

Researchers have introduced Optimal Transport Depth Up-Scaling (OT-DUS), a novel method for efficiently increasing the size of pre-trained large language models (LLMs). Unlike existing techniques that copy or average layers, OT-DUS uses Optimal Transport theory to align and fuse functionally corresponding neurons. This approach aims to mitigate performance degradation caused by neuron permutation mismatches. Experiments show OT-DUS outperforms current methods in both general and specialized domains, with performance gains increasing when new layers are inserted higher in the model architecture. AI

IMPACT Offers a more efficient way to scale LLMs, potentially reducing training costs and improving performance.

RANK_REASON Academic paper detailing a new method for LLM training. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• alphaXiv
• arXiv
• CatalyzeX
• DagsHub
• Gotit.pub
• Hugging Face
• large-language models
• Mingzi Cao
• Optimal Transport
• Optimal Transport Depth Up-Scaling
• ScienceCast

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →