Reproduce “Controlling and Assessing Appropriate Persona Use in LLM-based Dialogue Generation” 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
Jongkyung Shin, Inkyu Lee, Chiehyeon Lim
SCONPOS reduces irrelevant persona use by comparing a model’s hidden representations for prompts containing all persona attributes and only topic-matched attributes, then subtracting their difference during generation. On Llama3.1 PersonaChat, it lowers Overuse from 0.917 to 0.743 and raises PAS from 0.049 to 0.301, while ROUGE-L rises from 0.086 to 0.103. These benchmark results do not guarantee naturalness across conversations.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.