Reproduce “Balancing Stylization and Truth via Disentangled Representation Steering” 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
Chenglei Shen, Zhongxiang Sun, Teng Shi, Xiao Zhang, Jun Xu
StyliTruth, developed by researchers at Renmin University of China, is a training-free framework that preserves truthfulness in Large Language Models (LLMs) while enabling stylistic transfer through disentangled representation steering. This method improved a combined style-truthfulness metric by up to 30.65% over existing baselines.
Key insights
- Stylization-induced truthfulness collapse occurs because style-relevant and truth-relevant representations are entangled within specific attention heads, causing steering vectors to interfere.
- Different attention heads and layers within LLMs specialize in encoding either style or truth, with truth sensitivity peaking in intermediate layers.
- Disentangling style and truth representations into mutually orthogonal subspaces is feasible and crucial for enabling independent control without degrading factual accuracy.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.