Reproduce “LoFiT: Localized Fine-tuning on LLM Representations” 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
Fangcong Yin, Xi Ye, Greg Durrett
LOFIT introduces a two-step framework for adapting large language models by first identifying task-relevant attention heads and then fine-tuning minimal offset vectors. The method achieves performance comparable to state-of-the-art PEFT techniques while utilizing 20x to 200x fewer trainable parameters and exhibiting improved out-of-domain generalization.
Key insights
- Learning-based intervention strategies (fine-tuning) are more effective for modifying LLM representations than learning-free methods, even with limited data.
- Explicitly localizing and targeting a sparse, task-specific subset of attention heads is crucial for efficient and effective LLM adaptation.
- The identified attention heads responsible for different tasks (e.g., truthfulness vs. factual recall) are largely distinct and non-interchangeable.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.