Reproduce “BQ-LoRA: Binary-Quantized Low-Rank Adapters as Implicit Regularizers for Parameter-Efficient 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

Juyoung Park, Heejun Lee

BQ-LoRA targets adapter memory in multi-adapter serving by replacing QLoRA’s FP16 low-rank updates with binary signs and a learned per-layer scale initialized to zero. Training retains full-precision latent weights, so its peak memory matches QLoRA’s. On Qwen2.5-7B, LearnThresh uses 1,344 KB rather than 19,712 KB and scores higher on Avg*, but results do not establish wins across models or tasks.

Summary by alphaXiv.

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