Reproduce “The Persona Hierarchy Model: Understanding Contextual Generalization in Fine-Tuning 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
Jiachen Zhao, Zhengxuan Wu, David Bau, Weiyan Shi
The authors test how fixed system prompts affect transfer of fine-tuned behaviors. Across four behaviors and 15 contexts, pre-training distance between a prompt’s persona vector and the default correlated with narrower transfer, though strength varied by behavior. In coding reinforcement learning, PPR reduced reward hacking from 53.6% to 0% under the Anti-hack prompt; this finding is limited to that setup.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.