Reproduce “StructLens: A Structural Lens for Language Models via Maximum Spanning Trees” 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

Haruki Sakajo, Frederikus Hudi, Yusuke Sakai, Hidetaka Kamigaito, Taro Watanabe

StructLens provides a way to inspect how a model organizes token representations within layers. It compares residual-stream representations using Euclidean distance and links each token to its most similar predecessor, forming an order-constrained tree. Consecutive-token subtrees are more common in middle layers and less common under shuffled order; training changes vary by subtree size and do not establish task improvement.

Summary by alphaXiv.

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