Unveiling Multilingual LLM Structure: Cross-Layer Transcoder Approach

This research explores the inner workings of multilingual Large Language Models (LLMs), focusing on the representation of different languages across layers. The use of cross-layer transcoders offers a novel perspective on how these models process and integrate multilingual information.

• •Investigates the internal representations of multiple languages within LLMs.
• •Employs cross-layer transcoders for analyzing language mappings.
• •Contributes to understanding of how LLMs handle multilingual data.