Reproduce “No Refund for the Token Tax: Speculative Decoding Across 21 Languages” 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
Hanyu Yang
The paper tests whether built-in speculative drafters offset extra multilingual translation tokens. Qwen3.5-9B used 4.7 times as many tokens for Khmer as English across 200 FLORES-200 sentences. Replay tests covered 21 languages with Qwen3.5-9B and Gemma 4 E4B-it; English had the highest acceptance length in both families in these translation tests, not a general wall-clock speedup.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.