Adaptive level modification via player skill classification and large language models

Clinical Snapshot

Key Findings

• P Value: Not reported — no inferential statistical testing conducted

• Effect Size: Full-level playability: 5.9 percentage point reduction below baseline (74.1% vs 80.0%). Chunk-level playability: 3.5 percentage point improvement above baseline (83.5% vs 80.0%). Classifier accuracy: 97.82% on the reported test set.

• Primary Outcome: Post-modification level playability rate: 74.1% at full-level granularity and 83.5% at isolated-chunk granularity, compared to an 80.0% baseline playability rate of original unmodified Super Mario Bros. levels

• Nnt Or Sensitivity: Not applicable in the clinical sense. Classifier performance: 97.82% overall accuracy. Sensitivity and specificity per skill class (expert, normal, beginner) are not individually reported. Playability failure rate at full-level granularity: 25.9% (1 in approximately 4 modified levels is unplayable).

• Confidence Interval: Not reported for any outcome measure

Clinical Application

The framework's feasibility for real-world deployment is uncertain. Real-time LLM inference introduces latency that may be incompatible with live gameplay requirements. The 25.9% full-level unplayability rate would be unacceptable in clinical or commercial deployment without further refinement. The reliance on proprietary LLM APIs raises cost, privacy, and reliability concerns for healthcare applications. The framework has only been validated in a single, well-defined 2D platformer environment. This study has no direct relevance to PBS listings, TGA regulatory pathways, or RACGP clinical guidelines in its current form. Indirect relevance exists within the Australian digital health and medical device regulatory landscape: the TGA's Software as a Medical Device (SaMD) framework would apply if such adaptive gaming technology were deployed for therapeutic purposes (e.g., cognitive rehabilitation, paediatric occupational therapy). The Australian Institute of Health and Welfare (AIHW) has identified digital health engagement as a priority area. Any clinical deployment in Australia would require TGA SaMD classification assessment, ethics committee approval for human participant data collection, and compliance with the Australian Privacy Act 1988 regarding continuous behavioural monitoring. RACGP guidelines on digital health tools for chronic disease management may become relevant if adaptive gaming is integrated into care plans. This is not a clinical study and has no direct patient population. Potential indirect relevance exists for digital health applications including: gamified cognitive rehabilitation platforms, paediatric therapeutic gaming, motor rehabilitation using video game interfaces, and mental health engagement tools. In these contexts, adaptive difficulty is clinically relevant for maintaining therapeutic engagement and avoiding frustration or boredom in patient populations.

Abstract

Maintaining player engagement in video games requires a careful balance between challenge and player competence. Static difficulty settings fail to account for individual skill variation, while existing dynamic difficulty adjustment systems are limited to tuning low-level game parameters rather than restructuring level content. This paper presents an adaptive level modification framework that personalizes gameplay by continuously inferring player skill and applying targeted structural modifications to level content in real-time. A hybrid behavioral dataset is constructed by combining agent-generated trajectories, produced by Proximal Policy Optimization (PPO) agents, a reinforcement learning approach, trained at three distinct skill levels, with manually collected human gameplay data labeled through clustering. A classifier trained on this dataset categorizes players into expert, normal, and beginner skill levels, achieving an overall accuracy of 97.82%. The classifier output drives a two-stage large language model (LLM) pipeline guided by prompt engineering, which expands a skill-conditioned prompt into a structured modification instruction applied to the current level chunk. A physics-constrained verifier based on a graph-based shortest path method ensures all modified levels remain traversable. Evaluated on Super Mario Bros. levels, the framework achieves a post-modification playability rate of 74.1% at the full-level granularity and 83.5% at the isolated-chunk granularity, closely matching the 80.0% baseline of the original levels.

References

1. 1.Elshamy, A. A., Aliedin, H. N., Shaban, S. T., Aldakar, M. H., & Zaky, A. B. (2026). Adaptive level modification via player skill classification and large language models. Scientific Reports. https://doi.org/10.1016/J.ASOC.2019.105909