Reproduce “Emergent Structure in the Marginal Attention Space of 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

Valentino Maiorca, Walter Nelson, Francesco Locatello

The authors define marginal attention by averaging value-norm-weighted attention over output positions, then summing a head-by-token matrix to obtain token and head profiles. Across 67 open models, token profiles agree on the same text (mean Spearman rho 0.87), while head profiles are stable within models (rho 0.99 across document halves). At 8x compression, Qwen3-8B scores 91.8 with the head budget versus 72.8 with AdaKV and the same token scorer; latency gains are not established.

Summary by alphaXiv.

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