Microsoft released exciting research on a method that can improve even existing LLMs, which they have called Layer-Selective Rank Reduction (LASER).
They are seeing 20-30% model improvements with reduced memory requirements.
It appears relatively easy (for the engineers) to implement, so I believe we’ll hear more about LASER as it’s applied in the real world.
https://www.theverge.com/2024/1/31/24057362/microsoft-llm-accuracy-laser-research-ai
0 Comments