Reproduce “PersonalBench: Measuring the Authorship Gap in LLM Personalization” 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
Yash Ganpat Sawant
PERSONALBENCH introduces a novel evaluation framework, grounded in authorship science, to assess whether LLM personalization methods truly alter the model's writing style rather than just content. The research demonstrates that current inference-time techniques fail to bridge a quantifiable "authorship gap," with LLM-generated text consistently falling below the stylistic fidelity of human writing.
Key insights
- Current inference-time personalization methods are insufficient to close the "authorship gap," as LLM-generated texts, even when personalized, remain stylistically distinct from human writing and fall below the human cross-author stylistic floor.
- LLMs exhibit a persistent, inherent authorship fingerprint or "generated-text regime" that prompt engineering alone struggles to fundamentally alter, dominating over attempts at style imitation.
- Different personalization evaluation metrics can yield significantly divergent conclusions and show near-zero correlation, indicating that "personalization" is a multi-faceted concept requiring comprehensive, multi-metric assessment to avoid misleading interpretations.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.