A BEHAVIOR-AWARE CONCEPTUAL FRAMEWORK FOR PREDICTING NEGATIVE HUMAN BEHAVIOR USING LARGE LANGUAGE MODELS

Negative affect is a contentious issue of affective computing and computer-assisted mental health, especially in a dialogical context, where emotions change over time. More context-sensitive interpretations of human behavior are possible with recent developments of Large Language Models (LLMs), but most of the current systems use a single utterance to interpret an emotion and do not provide longitudinal models which are based on psychologically grounded assumptions. In this article, the study will advance a behavior-sensitive neural framework of negative affect prediction with the help of LLMs based on multi-turn and longitudinal conversational analysis instead of single-turn and single-classified emotion. The suggested framework combines the representation of contextual language, psycholinguistic signs, and the tracking of behavioral trajectory to make the inferences about the trends in emotions and to clearly place the system in the frame of non-diagnostic and supporting. The framework also includes strategies of safety constrained responses, human ethics, and human-inthe-loop considerations in the deployment of the framework in emotionally sensitive applications. Representative results of the past benchmark studies on the conversational emotion datasets are discussed as the motivation to the design choices and as an explanation of the benefits of the contextual transformer-based architectures compared with the baseline models. Instead of a fully-implemented system, this work provides the design advice and conceptual insights based on the recent (1000, 2023,2025) literature on affective computing, empathetic AI, and digital mental health. The proposed framework will be used in future research to inform on behavior-aware, ethically responsible conversational agents, which can endorse emotional awareness and resilience without the involvement of clinical diagnosis and treatment.