Reproduce “TopoSteer: Topological Layer Selection for Activation Steering in LLMs” 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
Tuc Nguyen, Marcos Gutiérrez-del-Pozo, Eduardo Paluzo-Hidalgo, Yifan Hu, Adaku Uchendu, Thai Le
TopoSteer selects layers for paired-example mean-difference steering by measuring the connectivity of standardized activation displacements: it sums the minimum-spanning-tree edge lengths and ranks layers by that score, while steering uses raw mean-difference vectors. On Qwen2.5-3B-Instruct, its unit-sphere variant reaches 79.71% aligned-answer probability averaged across six tasks, versus LayerNavigator’s 66.20% for three simultaneous layers; the standardized variant scores 61.07%.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.