Reproduce “SYNC: A Training-Free Inference Controller for Dynamic Layer Selection in Multi-Persona Alignment” 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
Yosef Nuraga Wicaksana, Muhammad Reza Qorib, Fadhil Muhammad, Alfan Farizki Wicaksono
SYNC steers multiple behaviors by checking a model’s internal activations during generation and routing each behavior’s separate steering vector to a currently receptive layer. It adapts layer assignments during decoding, while merging vectors during prompt prefill. Across 28 persona configurations, SYNC led in 21 and significantly beat both baselines in 15; results cover two models.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.