Reproduce “Mixture of Convolutions: Token-Adaptive Short Convolution for Sequence Modeling” 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
Andrew Mouldon
MoC makes short causal convolution token-dependent by mixing learned local filters according to each token, while retaining the same short window. The authors apply it separately to projected queries, keys, and values. Across evaluated Transformer and Gated DeltaNet settings, MoC lowers FineWeb-Edu and WikiText-103 perplexity than static convolution; at 460M parameters, FineWeb-Edu perplexity is 13.67 versus 13.82, within tests up to 15B training tokens.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.