Reproduce “Low-Discrepancy Dither for Quantized Recurrent State Caches” 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

Snigdha Chandan Khilar

The authors study deterministic rounding for recurrent caches in Mamba-style models, where errors feed into later state updates. Weyl dither replaces fresh random thresholds with an evenly spread sequence. Across 15 tested settings, it reduced teacher-forced prediction divergence versus stochastic rounding in all settings, with a median reduction of 31%; this does not establish better generated-text quality or throughput.

Summary by alphaXiv.

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