Reproduce “Ensembles of Low-Rank Expert Adapters” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yinghao Li, Vianne Gao, Chao Zhang, MohamadAli Torkamani
Ensembles of Low-Rank Expert Adapters (ELREA) introduces a framework for fine-tuning Large Language Models by clustering training data based on instruction gradient similarity and training specialized LoRA experts on these clusters. This approach addresses conflicting gradient directions during optimization, leading to performance improvements of up to 9.67% over standard LoRA on mathematical reasoning tasks and demonstrating robust generalization on diverse language understanding benchmarks.
Key insights
- Gradient directions derived from instruction tokens effectively capture underlying semantic and structural similarities in diverse training data, enabling robust clustering.
- Clustering training data based on these gradient similarities allows for the creation of specialized LoRA expert adapters that mitigate conflicting gradients during fine-tuning.
- Dynamically weighting these specialized experts during inference based on input gradient similarity improves generalization and specialization without requiring explicit task-specific routing mechanisms.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.