YNTP-100: A Benchmark for Your Next Token Prediction with 100 People

Abstract

Large language models (LLMs) trained for general \textit{next-token prediction} often fail to generate responses that reflect how specific individuals communicate. Progress on personalized alignment is further limited by the difficulty of collecting real-world personal communication data due to privacy constraints. We propose Your Next Token Prediction (YNTP), a task that formulates personalized response generation as token-level prediction conditioned on user interaction history. We introduce \textbf{YNTP-100}, a benchmark built from multilingual multi-day human--agent conversations with 100 people, enabling systematic evaluation of user-specific response behavior. We evaluate external (parameter-preserving) and internal (parameter-updating) alignment methods using metrics of substance similarity and stylistic consistency. The dataset and results are publicly available at: https://github.com/AnonymousHub4Submissions/YNTP100.