Know You Before You Speak: User-State Modeling for LLM Personalization in Multi-Turn Conversation

A research team has proposed a framework called PUMA that aims to personalize large language model conversations by inferring hidden user states rather than relying solely on explicit user history, according to a paper submitted in 2026 [1]. The framework, detailed in a paper submitted to arXiv on 23 May 2026, is grounded in the Free Energy Principle (FEP) and formulates personalization as decision-making under partial observability [1][2]. At each turn in a multi-turn conversation, PUMA maintains a belief over the user's hidden state and selects dialogue actions by minimizing expected free energy, balancing what the researchers term epistemic and pragmatic objectives [2]. This approach shifts personalization from passive memory retrieval to model-based decision-making over user evolution [2]. The researchers note that existing memory- and profile-based methods primarily reuse observable user information, offering limited support for modeling user-state dynamics or selecting actions based on how they shape future user states [2]. PUMA was instantiated on healthcare-oriented counseling and motivational interviewing benchmarks with latent state annotations for evaluation [2]. Experiments showed that PUMA improved long-horizon dialogue outcomes while maintaining strong response quality, and a cross-dataset study demonstrated more reliable user-state estimation and next-state prediction [2]. The work arrives amid broader efforts to refine how generative AI models handle context. Prompt engineering, the practice of designing and refining input instructions given to a generative AI model to produce more accurate or useful outputs, has become a recognized skill for working with large language models [3]. Context engineering, a related area, focuses on the management of non-prompt contexts such as metadata and tokens [3]. PUMA represents a different direction, embedding user-state inference directly into the model's decision process rather than relying on externally engineered prompts. ChatGPT, the generative AI chatbot developed by OpenAI and originally released in November 2022, accelerated the AI boom and reached 900 million weekly active users by February 2026 [4]. The chatbot has been criticized for limitations including hallucinations—plausible-sounding but incorrect answers—and biases reflected in its training data [4]. Personalization frameworks like PUMA target a different challenge: making interactions more adaptive to individual users over extended dialogues without requiring explicit user profiles. Artificial intelligence has been applied across industry and academia for tasks including language translation, image recognition, and decision-making [5]. The subfield of machine learning has seen massive advancements in generative AI, which uses generative models to produce text, images, and other data [5]. The PUMA framework contributes to this landscape by proposing a method for language models to infer and act upon latent user states during conversation.