Personalizing Large Language Models (LLMs) is essential for user engagement. However, practical personalization often encounters a "cold-start" scenario where the interaction history is extremely limited. Existing methods typically rely on extensive pre-collected user data or simulated interactions, which are often unavailable or insufficient for real-world applications.
To solve the problems, a joint research team from Harbin Institute of Technology and Tsinghua University published their new research on 15 September 2026 in Frontiers of Computer Science co-published by Higher Education Press and Springer Nature.
The team introduced ColdChat, the first benchmark designed to assess LLM personalization using brief, real-world interaction histories. Furthermore, they proposed EPIC (Extracting user Profile from Interaction Context), a novel framework designed to extract user profiles dynamically during early interactions.
The methodology involved collecting multi-session dialogues from 101 real-world annotators to construct ColdChat. The proposed EPIC framework employs a two-stage architecture: a classifier-based "Judger" to identify personalization signals and an LLM-based "Generator" to integrate these signals into a coherent user profile.
The experimental data shows that compared with baseline methods that rely solely on dialogue history, the profiles extracted by EPIC yield substantial improvements, achieving a +27.2% gain in the personalized ranking task and a +12.7% gain in the personalized generation task.
This work establishes a foundational benchmark and a robust framework to advance LLM personalization, enabling chatbots to effectively understand and adapt to users from the very first interaction.
DOI:10.1007/s11704-026-51716-6