Steering Emotional Dynamics for Art Therapy: Controllable Narrative Script Generation through Hierarchically Guided LLM Agents

Researchers have proposed a new framework, EC-Script, designed to generate narratives for art therapy with controlled emotional dynamics, according to a paper submitted to arXiv on 15 Jun 2026 [1]. The system uses large language models to create scripts that adhere to a preset emotional trajectory, a capability existing methods have struggled to achieve [1]. The framework, detailed in a preprint on the open-access repository arXiv, addresses a core challenge in art therapy, where narrative creation is a primary vehicle for emotional expression [1][2]. The inherently dynamic nature of emotions during healing requires narratives with finely controlled emotional fluctuations, allowing individuals to safely project inner conflicts and achieve emotional catharsis [2]. While large language models (LLMs) have enabled automated narrative generation, prior systems could produce fluent text but failed to reliably follow a specified affective trajectory [2]. LLMs are machine learning models with many parameters, trained on vast amounts of text for language generation tasks [8]. EC-Script operates through a three-stage hierarchical process. It first establishes an overall narrative direction via Emotion-Trajectory Planning, then propels scene-level plot development with Character-Driven Scene Generation, and finally regulates local emotional changes through Emotion-Controlled Script Writing [2]. The output is scene-by-scene script content that remains highly consistent with the preset affective trajectory [2]. Experimental results demonstrated that EC-Script significantly outperformed baseline methods in affective trajectory adherence, exhibiting what the authors describe as excellent and reliable emotional controllability [2]. The paper was posted on arXiv, a repository of electronic preprints that is moderated but not peer-reviewed, which has hosted over two million articles since its launch in 1991 [6]. The submission appears alongside experimental community tools developed under the arXivLabs framework, a program that allows collaborators to build features directly on the site while adhering to values of openness, community, excellence, and user data privacy [4]. The arXivLabs initiative, formalized in 2020, provides a space for community innovation and includes tools such as the Bibliographic Explorer and CORE Recommender [4][5]. The arXiv team is currently pausing new Labs proposals while it focuses on modernizing and moving its systems to the cloud [3].