Reproduce “MoBiQuant: Mixture-of-Bits Quantization for Token-Adaptive Any-Precision LLM” 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
Dongwei Wang, Jinhee Kim, Seokho Han, Denis Gudovskiy, Yohei Nakata, Tomoyuki Okuno, KhayTze Peong, Kang Eun Jeon and 3 more
MoBiQuant lets a language model use weight precisions across tokens, using recursively quantized weight slices and a learned router to assign refinements under an average-bit budget. This targets outlier migration, where tokens sensitive to quantization change across bit-widths. On WikiText2, LLaMA3.2-3B at equal average 4-bit precision reaches perplexity 9.41 versus 12.71 with OmniQuant; results do not establish universal superiority.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.