Reproduce “LoopSkim: Lossless Multi-Depth Speculative Decoding for Looped 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
Muhan Xiao, Shijian Li
LoopSkim accelerates decoding in looped language models using intermediate readouts to check proposed tokens during existing computation. Its untrained context-copying method corrects mismatches before full depth; only full-depth readouts commit output, without a separate verifier. Across evaluated cells, it achieves a 1.73× geometric-mean ratio of speedups over LoopSpec, but uses 4.74–7.74 times as many loops per output token; identical output is guaranteed only in exact arithmetic.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.