Reproduce “LoRA-PAR: A Flexible Dual-System LoRA Partitioning Approach to Efficient LLM Fine-Tuning” 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

Yining Huang, Bin Li, Keke Tang, Meilian Chen

LoRA-PAR introduces a dual-system LoRA partitioning method, inspired by human fast and slow thinking, to efficiently fine-tune large language models. This approach attains competitive or superior performance on complex reasoning tasks, such as increasing GSM8K accuracy by approximately 12% over vanilla PiSSA, while utilizing a substantially reduced set of active parameters.

Key insights

  • Cognitively-inspired dual-process theory can guide LLM fine-tuning by specializing parameter subregions for distinct reasoning types.
  • Activating a carefully selected, importance-ranked subset of LoRA parameters is critical for efficiency and performance, outperforming random or full activation.
  • An SFT-then-RL training sequence effectively builds LLM capabilities, establishing intuitive understanding before refining deliberate reasoning.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.