Reproduce “TempQuant: Learning Adaptive Temperatures for Quantized Language Models” 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

Mukund Athmaram, Liang Ma, Manoj Rao

The authors introduce TempQuant, which tunes temperature and top-p sampling settings to make a quantized language model’s next-token probabilities better match its full-precision model, without changing model weights. Quantization spreads probabilities but rare top-token ranking flips account for the mismatch. On MATH, global TempQuant raises 3-bit Qwen3-4B accuracy from 33.8% to 51.0%; sampling cannot reorder tokens or guarantee recovery.

Summary by alphaXiv.

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