LETGAMES: An LLM-Powered Gamified Approach to Cognitive Training for Patients with Cognitive Impairment

Jingwei Shi, Shengyu Tao, Xinxiang Yin, Chen Huang, Wenqiang Lei, See-Kiong NgarXiv 2026 · DOI

TL;DR

This paper introduces LETGAMES, an LLM-powered system that automates the creation of personalized, safe, and engaging gamified experiences for cognitive training in patients with cognitive impairments. By leveraging a multi-agent architecture and a novel evaluation protocol, LETGAMES addresses the labor-intensive nature of traditional serious game development, offering a scalable solution for accessible cognitive rehabilitation.

What Problem Does This Paper Address?

Cognitive impairment, exacerbated by an aging population and neurodegenerative diseases, presents a growing public health challenge. Cognitive training, particularly through Serious Games (SGs), offers a promising non-pharmacological intervention. However, developing effective SGs is a complex, resource-intensive process requiring collaboration between game designers and clinical therapists. This manual approach leads to high costs, slow development, and often results in "one-size-fits-all" games that fail to personalize content to individual cognitive deficits, symptom profiles, or psychological needs. Furthermore, ensuring game safety and mitigating negative emotional responses in vulnerable patients is paramount but often overlooked. This paper tackles the challenge of creating personalized, safe, and effective cognitive training SGs at scale by automating the design process using LLMs.

Methods

LETGAMES employs a dual-track multi-agent architecture: a Game Master (GM) and a Psychology Master (PM).

  • Game Master (GM): This component is responsible for generating personalized game scenarios and challenges. It comprises:
  • Game Designer (AGD): Creates game specifications (scenarios and tasks) tailored to a target cognitive domain and patient profile. It uses a slot-filling paradigm and predefined schemas, inspired by "Dungeons & Dragons," to generate open-world interactive narratives. For memory and verbal learning, it employs an "Encoding-Retention-Retrieval" strategy.
  • Game Controller (AGC): Acts as the runtime engine, managing game state, ensuring narrative coherence, and generating immediate gameplay challenges. It uses a memory module to maintain continuity and adapts to player actions.
  • Game Critic (AGCr): Evaluates game content for quality, narrative coherence, factual accuracy, and clinical boundaries, issuing revision requests to AGC.
  • Psychology Master (PM): This component ensures psychological safety and provides support. It includes:
  • Hint Provider (AHP): Offers tiered scaffolding instructions (gentle nudge, strategic guidance, direct instruction) to prevent frustration and guide players.
  • Emotion Copilot (AEC): Monitors player emotional states and intervenes with conversational strategies to mitigate distress, offering relaxation techniques or transitioning to simpler scenarios.

LETGAMESEVAL, a novel psychology-grounded evaluation protocol, is also introduced. It uses therapeutic interaction analysis and a comprehensive metric suite to assess therapeutic quality, safety, game content, and user engagement, avoiding resource-intensive clinical trials.

Datasets

The paper mentions using the D4 dataset (Yao et al., 2022) for generating patient profiles, including demographic variables and depression symptoms. These profiles are then augmented with synthetically generated life experiences by LLMs.

Key Insights and Ideas

  • LLMs as Automated Game Designers: The core innovation is using LLMs to automate the complex and labor-intensive process of designing personalized serious games for cognitive training, moving beyond manual expert design.
  • Multi-Agent Architecture for Holistic Support: The separation of GM and PM roles allows for simultaneous optimization of therapeutic efficacy (gameplay) and psychological safety (emotional support), addressing a critical gap in existing SGs.
  • Adaptive Difficulty and Personalized Scenarios: LETGAMES dynamically adjusts game difficulty based on player performance and tailors scenarios to individual patient profiles and life experiences, enhancing engagement and effectiveness.
  • Psychology-Grounded Evaluation: The LETGAMESEVAL protocol provides a robust, scalable, and ethically sound method for evaluating SGs, focusing on therapeutic interaction analysis rather than solely on direct cognitive outcome measures.
  • Importance of Emotional Support: The AEC's role highlights the critical need to manage patient emotions and prevent frustration, which can significantly impact therapeutic progress and self-efficacy.

Key Results

LETGAMES demonstrated significant superiority over the baseline (ReMe) across nearly all evaluated metrics, including therapeutic efficacy, safety, and engagement.

  • Therapeutic Training Quality: LETGAMES consistently achieved higher scores in Helpfulness and Domain Alignment compared to ReMe across various user groups (LLM-based SPs, LLM-based Controls, Human-portrayed SPs, Human-based Controls) and LLM backbones (GPT-4o Mini, GPT-4o).
  • Game Safety: LETGAMES achieved near-perfect scores in Safety, Necessary Hints, and Anxiety-free interactions, indicating robust content and effective emotional support.
  • Game Content & User Engagement: LETGAMES also outperformed ReMe in Coherence, Personalization, Enjoyment, and Willingness, suggesting more engaging and tailored experiences.
  • Adaptability: Domain-specific evaluations showed LETGAMES maintained consistent efficacy across different cognitive domains (Memory, Attention, Verbal Learning, Executive Function, Social Cognition) and age groups, highlighting its adaptability.

Limitations and Future Work

  • Generalization to Real Patients: The current evaluation relies on simulated patients (LLM-based and human-portrayed SPs). While this aligns with common practice, it may not fully capture the complex, heterogeneous nature of real-world cognitive impairment. Future work requires deploying LETGAMES in authentic clinical settings with diverse patient demographics.
  • Long-Term Treatment Efficacy: The current evaluation methodology assesses immediate effectiveness. The long-term therapeutic efficacy and persistence of cognitive gains over extended periods need to be validated through large-scale, longitudinal clinical trials.
  • Multi-Domain Training: Current frameworks often focus on isolated cognitive functions. Future research should prioritize developing SGs capable of integrated, multi-domain training to address complex clinical manifestations.
  • Multimodal Extensions: While LETGAMES is dialogue-driven, future work could explore integrating multimodal elements (e.g., visual representations, adaptive music) to enhance immersion, though this requires careful validation for safety and impact.

Summary updated

BibTeX
@misc{shi2026letgames,
  author = {Jingwei Shi and Shengyu Tao and Xinxiang Yin and Chen Huang and Wenqiang Lei and See-Kiong Ng},
  title = {{LETGAMES: An LLM-Powered Gamified Approach to Cognitive Training for Patients with Cognitive Impairment}},
  year = {2026},
  eprint = {2604.09566},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2604.09566},
}

Loading related content...

Figure