Researchers Found a Better Way to Teach Large Language Models New Skills
10 months ago
- #AI
- #Machine Learning
- #Natural Language Processing
- Researchers developed WeGeFT, a technique improving large language models' performance without additional computational power.
- WeGeFT enhances model performance in tasks like commonsense reasoning, arithmetic reasoning, and code generation.
- The technique builds on LoRA (2022) but adds mathematical tools to identify and prioritize novel parameters for learning.
- Proof-of-concept testing shows WeGeFT matches or outperforms LoRA and its variants across various tasks.
- Future work explores using WeGeFT to identify harmful outputs and improve AI alignment and safety.
- The research will be presented at the International Conference on Machine Learning in July 2024.