Reproduce “Latent Inter-User Difference Modeling for 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
Yilun Qiu, Tianhao Shi, Xiaoyan Zhao, Fengbin Zhu, Yang Zhang, Fuli Feng
The Difference-aware Embedding-based Personalization (DEP) framework advances LLM personalization by modeling inter-user differences in a latent space, distilling these signals using a sparse autoencoder. It consistently outperforms state-of-the-art methods on personalized review generation tasks, yielding significant improvements in metrics like ROUGE-1, METEOR, BLEU, and BERTScore.
Key insights
- Modeling inter-user differences in a continuous latent embedding space is more effective and robust for LLM personalization than relying on natural language prompts.
- Sparsity-constrained distillation via a Sparse Autoencoder is crucial for refining personalized embeddings, filtering noise, and retaining only the most salient, task-relevant features.
- The combination of user-specific historical context and explicit difference-aware signals in the latent space provides complementary information, leading to superior personalization.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.