In multiplayer online games, intelligent matchmaking systems are the core scheduling mechanisms responsible for dynamically balancing player demands and system states. However, the traditional rule-based matchmaking system tries to balance matching quality and time consumption by manually adjusting multiple preset parameters, which often has poor real-time adaptability in response to the dynamic fluctuations of players. To make a smarter system, a “predict-then optimize” matchmaking (PTOM) framework is proposed, which integrates a Feature-interactive Behavior Sequential Transformer (FBST) prediction model and a Lagrangian dual-based dynamic programming algorithm. FBST uses a monotonicity constraint (ensuring that longer waiting times are associated with higher success rates) to predict the matching success rate based on real time player online data, while the optimization algorithm can use the prediction results to automatically obtain the best solution to multiple preset parameters through constructing and solving an optimization problem. Offline experiments demonstrate that FBST outperforms baselines in prediction accuracy and mono tonicity preservation, while the optimization algorithm achieves higher rewards and smaller duality gaps. A/B tests on online game (20 days, per-click randomization, reported per mode) show that the framework significantly improves the matching success rate (by 2.77% in normal and 0.93% in ranking matches) and reduces average matching time (by 20.40% in normal and 7.99% in ranking matches), validating its effectiveness in handling real time high-dimensional matchmaking challenges. We release the code at https://anonymous.4open.science/r/PTOM-E4DB/.
Cybersickness remains a major barrier to virtual reality adoption in serious games, yet audio-based mitigation is underexplored compared with visual interventions. This pa per presents an anticipatory audio architecture that predicts upcoming motion events and generates procedural sound cues before the visual motion occurs, implemented with MetaSounds in Unreal Engine 5. The design is motivated by predictive coding accounts of motion sickness, in which advance auditory information about impending motion may help the brain form more accurate sensory predictions. Technical evaluation on a single representative maritime scene shows motion-to-audio latency below 15 milliseconds with modest computational overhead. We report an exploratory cross-sectional study with 105 participants. Because all participants experienced the same anticipatory-audio configuration and no control condition was included, the study cannot establish the causal efficacy of anticipatory audio; it tests associations, not effects. Participants who perceived higher audio quality reported less motion sickness; the association was carried mainly by perceived audio–visual synchronization, survived demographic adjustment, and attenuated to marginal significance once session duration and cognitive load were controlled. Physiological measures corroborated the sickness self-report rather than the intervention. We conclude with design guidelines labelled as empirically supported or engineering-derived and outline the controlled comparison needed to establish efficacy.
Mass-casualty incident (MCI) training requires diverse scenarios while maintaining instructional consistency. However, most existing digital triage simulations remain script driven, relying on manually authored patient cases and predefined injury assets that limit scalability and replayability. This paper proposes a hybrid procedural serious game framework that combines LLM-based semantic patient synthesis with deterministic START validation to generate controllable training scenarios. Given predefined scenario specifications, structured patient profiles are procedurally generated and integrated into a two-layer architecture that separates semantic patient generation from geometric scene construction. To support safe-aware visual generation, validated injury descriptors are transformed into safety-constrained prompts for stylized low-poly visual generation while preserving triage-relevant cues. Experimental results demonstrate controllable variability across different seeds, along with repeatability of patient records and scene configurations under fixed conditions. Generation reliability did not improve consistently with model size. A four-week randomized study further shows that repeated gameplay improves triage decision speed relative to the control condition, while providing no significant additional gain in accuracy. These findings show that LLMs can function as controllable semantic generators within hybrid procedural frameworks, supporting replayable MCI training while retaining rule-based triage validation and safety aware visual design.
Large language models (LLMs) are increasingly used as autonomous agents in interactive environments, yet their behavior in competitive dialogue-driven games remains poorly explored. Text-based social-deduction games provide a controlled testbed because agents must use language to persuade, coordinate, deceive, and form coalitions under uncertainty. This paper introduces a reproducible protocol for studying competitive coalition formation among LLM agents. Three agents negotiate survival in a resource-scarce environment using private messages while maintaining explicit, inspectable relational state. Coalition closure is implemented through a discrete exclusion-commitment state, which represents protocol-level support for an exclusion decision. We evaluate 80 runs organized as a $2\times 2$ design crossing model composition and memory: homogeneous same-backend agents versus heterogeneous mixed-backend agents, each tested with and without moderator-mediated chain memory. Homogeneous agents reach agreement in all 40 runs and converge quickly. Heterogeneous agents reach agreement in 32 of 40 runs, require substantially more rounds and messages, and account for all no-agreement episodes. Memory has condition-dependent effects: it has limited impact in homogeneous play, while in heterogeneous play it is associated with higher agreement and slightly fewer rounds/messages, without removing the coordination cost of model heterogeneity. Agents spontaneously produce quasi-contractual bargaining language with no external enforcement.
Competitive online gaming remains male-dominated, and toxicity disproportionately targets non-male players. Under the veil of anonymity, players with marginalized gender identities encounter discrimination, hate speech, and unwanted sexual advances. Addressing these issues requires understanding how players perceive gender-based harassment. This work investigates gender discrimination in Competitive Online Multiplayer Games through two surveys. In the first survey, 61 non-male players assessed gender-based discrimination and reported their use of moderation tools, including identity concealment or disclosure. Their responses yielded 171 statements describing forms of gender discrimination. In the second survey, 281 players across all gender identities rated these statements for severity and authenticity and reported how they use in-game moderation tools in response to discriminatory behavior. Our results indicate that gender discrimination is perceived as similarly severe across gender groups, but responses differ: players from marginalized groups report higher levels of rumination. Higher perceived severity predicts increased moderation tool use; male players show a stronger rise in muting offenders, whereas female players show a stronger rise in reporting them. Among previously targeted players, perceived severity does not predict gender concealment. These findings offer guidance for game developers seeking to design more inclusive gaming environments.
This paper presents a longitudinal study examining whether game-based learning can improve music notation reading more effectively than traditional methods in real-world practice environments. Thirty adult keyboard learners completed pre- and post-assessments over a one-week period, practising with either The Crypt of Notation, a purpose-built educational game, or a widely-used online notation tool. Results demonstrate that the game group achieved over 30% greater improvement across all reading assessment criteria, including correct note identification and rhythm accuracy, while the control group showed no significant gains and measurably declined on the most complex material. Game group participants also practised beyond the minimum requirement, reporting significantly higher motivation and confidence throughout, while control group engagement declined steadily across sessions. These findings suggest that well-designed game-based approaches can simultaneously develop foundational notation skills and sustain the engagement necessary for long-term musical progression, directly addressing the dropout factors that disproportionately affect adult learners.
Modern game quests are increasingly trending toward high freedom and multiple endings, making their design and testing more complex. Smoke testing in games aims to rapidly validate quest functionality, but existing automated testing approaches face critical limitations when handling quests with branching narratives, and multiple possible endings. These challenges result in insufficient exploration of endings, redundant action sequences, and prohibitively high interaction costs. To address these challenges, we propose EndWeaver, an automated smoke testing framework specifically designed for multi-ending game tasks. EndWeaver leverages LLMs to interpret quest descriptions and game environments, generating high-level methods to guide global exploration, complemented by the generated primitive task sequences for fine-grained planning. By combining this multi-level planning structure with the dynamic construction of state-transition graphs, EndWeaver ensures comprehensive coverage of all possible endings and enables efficient testing. We evaluated EndWeaver on four game environments—Mini RPG, Overcooked Plus, Minecraft, and PySC2 against baseline methods. In all environments, EndWeaver achieved stable quest completion with high reliability, significantly outperforming the baselines. In terms of testing cost, EndWeaver required only a few thousand steps, while baseline methods required millions of steps, all while ensuring comprehensive coverage of all possible quest endings. Ablation studies further demonstrated the critical contributions of each framework component, validating the effectiveness and robustness of our approach.
Visualization is a key technique in game analytics, enabling analysts to interpret player behavior through intuitive graphical representations. While aggregated visualizations such as heatmaps and charts are widely used, they often obscure the sequential structure that is essential for analyzing player decision-making processes and problem-solving strategies. Existing process visualization systems either focus on identifying optimal paths or offer limited interactivity, typically restricted to basic filtering and inspection capabilities. To address this gap, we conducted an iterative design process with three professional game analysts across distinct game contexts, including an open world social entertainment game, an educational quest based RPG, and a Massively Multiplayer Online Role Playing Game (MMORPG). This process resulted in three key features: (1) Segmentation and the Difference Graph for partitioning players into subgroups based on metadata or behavioral patterns and visually comparing differences between these groups; (2) Backward Tracing for identifying common preceding paths to critical events and understanding player progression patterns; and (3) Pattern-based Re-abstraction for aggregating low-level actions into insightful higher-level constructs such as skill combos. A between-subjects utility study with 44 participants showed that analysts using the designed features achieved significantly higher accuracy in interpreting sequential player behavior. And a usability study with 12 analysts yielded a mean SUS score of 77.3, indicating good perceived usability. These results demonstrate that the proposed features address real world analytical needs, uncover insights overlooked by existing systems, and provide a transferable foundation for future process oriented visualization systems in game analytics.
Humor understanding remains a distinct challenge for large language models (LLMs), particularly in vertical domains rich in cultural nuances, such as video games. In this paper, we introduce Shumor, a bilingual benchmark for gaming humor understanding in Steam game reviews. By benchmarking state-of-the-art (SOTA) LLMs on Shumor for humor intensity prediction, we observe a gap between model predictions and the human-validated humor-intensity labels in gaming contexts. For the tasks of humor classification and explanation, existing models still show limitations in domain knowledge and structured reasoning within the gaming domain. Therefore, we propose SOPA-SC (Script-Opposition-Purpose Architecture via Synergistic Collaboration), a structured multi-perspective reasoning framework that combines chain-of-thought (CoT) reasoning with role-conditioned perspective synthesis. By organizing complementary Gamer, Linguist, and Psychologist perspectives, SOPA-SC decomposes complex domain-specific humor into game-contextual, rhetorical, and communicative-intent dimensions. Across the evaluated models, SOPA-SC increases the average F1-score and Jaccard Index for humor classification. Human and automatic evaluations also indicate that SOPA-SC generallyproduces more detailed and context-specific explanations than the baseline setting.
This paper introduces a novel framework for efficiently evaluating in-game sound designs using multimodal learning and reinforcement learning. Sound design significantly influences gameplay and accessibility; however, objective evaluation methods, defined here as automated assessments based on agent performance rather than human ratings, often require retraining agents from scratch for each sound design. Our approach first pre-trains a multimodal encoder that encodes visual and auditory information into a shared latent space. A reinforcement learning agent is then trained using this latent representation. We fine-tune the multimodal encoder for each sound design while keeping the trained policy fixed, allowing us to isolate the effects of sound changes on agent behavior. To evaluate the proposed method, we conducted experiments on games from multiple genres, comparing the performance of blind agents (agents using only sound input) under various sound designs. Experimental results suggest that this approach can detect differences in agent performance caused by degraded or altered sound design. Moreover, we significantly reduce the computational cost compared to retraining the agent from scratch. Experimental results suggest that this approach provides a practical and scalable solution for evaluating in-game sound designs. The source code, data and supplementary document are available https://github.com/thainv0212/tog-reusing-agents.
Serious video games have proven effective in multiple fields, although their design, development, and validation remain non-trivial, especially in domains requiring collaborative processes. In software engineering education, teaching the Scrum methodology is particularly challenging, as conventional instruction conveys terminology but rarely reproduces the teamwork of real-world environments. This paper presents SCRUMVERSE, a theory-driven, multiplayer virtual reality (VR) serious game designed to teach Scrum through embodied role-play, immediate feedback, and synchronous team coordination. Grounded in Kolb's Experiential Learning Cycle and Keller's Attention, Relevance, Confidence, and Satisfaction motivational model, the study examines the impact of the game on learning outcomes and student experience. A cluster randomized controlled trial was conducted with 220 students in software engineering, who played the game or completed a lecture-based control activity. Knowledge acquisition was measured through pre/post-tests, while post-intervention perceptions were captured using two validated questionnaires (GAMEX and Presence) and an ad hoc questionnaire assessing usability, learning, and motivation. Results show that both groups achieved significant learning gains, with the experimental group obtaining a statistically significant advantage. Students rated SCRUMVERSE highly for learning impact, instructional clarity, and presence, while reporting moderate enjoyment. Overall, SCRUMVERSE enhances cognitive and affective dimensions of learning by enabling authentic collaborative experiences aligned with Scrum principles.
Automated game testing is essential for ensuring software quality and player experience, yet conventional methods struggle to cover intricate interactions and uncover deep-seated defects. We propose PBSEGTE, a novel paradigm that integrates defect-oriented deep reinforcement learning with program-guided player-behavior simulation to enhance exploratory game testing. Our framework unifies visual perception, behavior modeling, and adaptive strategy switching. We design a spatio–temporal attention proximal policy optimization algorithm that learns diverse behavior patterns from real-player telemetry via heterogeneous behavior fusion modeling. A multidimensional defect-reward mechanism, incorporating visual anomaly detection and code-coverage analysis, steers the agent toward high-risk code paths. A Bayesian adaptive decision-tree controller dynamically switches among exploratory strategies based on coverage feedback. Extensive experiments in complex open-world environments demonstrate that our approach significantly outperforms random testing, record-and-replay, and the state-of-the-art DeepMind STAC baseline, achieving a logic-error detection rate of 42.7% and path coverage of 79.3%. This study establishes a solid foundation for efficient, self-adaptive automated game testing.
This paper presents beat understanding (BU), a deep learning-based framework for automatic chart generation in the rhythm game Taiko no Tatsujin. We propose a comprehensive preprocessing pipeline that incorporates annotation normalization, where the annotations refer to manually labeled hit events in TJA files, together with sequence length alignment to enhance training stability and output quality. We developed preprocessing techniques that incorporate beats per minute alignment, rhythm discretization, and sequence padding, together with standardized Mel spectrogram conversion and annotation formats, to eliminate inconsistencies in the raw data. Additionally, a novel audio segmentation strategy is applied in conjunction with a temporal error correction algorithm, significantly improving model prediction accuracy. Experimental results show that the BU framework demonstrates strong performance in both note prediction and human-like rhythm pattern generation. The proposed method improves adjacent detection accuracy by approximately 12.49% and localized unique-pattern overlap by approximately 29.73% compared with TaikoNation. At the same time, its overall performance shows a high degree of similarity to human-authored charts. These findings validate the effectiveness of the proposed slicing and alignment strategies in supporting structural modeling. Overall, this study provides empirical evidence and methodological guidance for advancing automatic chart generation in rhythm-based games.
The development of autonomous systems for high-stakes environments requires simulation frameworks capable of modeling complex multi-agent interactions and facilitating rigorous research. In this paper, we introduce NavySim 2.0, an integrated, algorithm-agnostic framework and simulation engine. It serves as a modular testbed for developing, integrating, and evaluating diverse maritime algorithms for maritime research and education. By mimicking real-world conditions within a controlled and risk-free environment, the platform enables researchers to rigorously train and validate autonomous algorithms while also assisting seafarers in mastering essential maritime skills, such as ship maneuverability, collision prevention, and understanding other naval agents' intentions. We present an implementation of a Unity-based naval simulator that provides the following key capabilities: 1) the ability to create and save complex navigation scenarios, with multiple vessels equipped with vulnerability heatmaps to represent their capabilities, 2) COLREGs collision avoidance for all vessel traffic, 3) detection of threats from surrounding traffic, based on a threat heatmap that encapsulates Closest Point of Approach (CPA) information, 4) identification of vessel intentions based on Hidden Markov Models (HMMs), 5) dynamic responses to perceived threats, and 6) visualization of threats and responses in the simulator. This paper describes the naval simulator with its capabilities and illustrates its main capabilities in various naval scenarios.
Generating accurate parametric human models from multi-cue observations is a fundamental challenge in producing high-fidelity game characters. Parametric reconstruction is essential for digital character production, animation, and interactive physics simulation within game environments. Existing alignment methods remain unstable under complex clothing, extreme poses, or severe occlusions. The instability severely degrades the generation of high-quality game ready characters. To tackle the challenge, we present a robust character reconstruction framework to systematically resolve the geometric discrepancy between clothing and human skin. Specifically, we propose a tightness-aware mechanism to filter geometric noise from loose garments through localized contact analysis. We subsequently introduce an implicit dense correspondence field to handle incomplete observations. The implicit field establishes a reliable mapping between unstructured inputs and the canonical model. Finally, we design a decoupled training strategy to optimize the undressing and character modeling stages independently. The decoupled approach enables the network to effectively handle diverse outfits. Extensive experiments demonstrate that our proposed model provides a reliable and highly accurate solution for character creation in game development pipelines.
This paper presents a 10-year systematic review (2015-2025) of the interface architectures and interaction modalities of serious games designed for children with Autism Spectrum Disorder (ASD). The analysis shifts the focus from general therapeutic efficacy toward the engineering of user interfaces, evaluating how different paradigms—graphical, tangible, motion-based, and AI-supported—impact the system's ability to meet neurodiverse requirements. A total of 15 empirical studies were analyzed across various interface modalities, including Graphical User Interfaces (GUI), Tangible User Interfaces (TUI), Mixed Reality (MR), and motion-based or Virtual Reality (VR) systems. Results show a shift toward embodied and multimodal interaction models, highlighting advances in sensor integration, feedback design, and adaptive control mechanisms. Tangible and hybrid systems suggested superior engagement and social initiation outcomes within the analyzed cohorts, while MR-based platforms showed potential for emotion recognition. However, technical challenges persist in scalability, real-time responsiveness, and maintenance of assistive technologies. The paper concludes with a proposed taxonomy of interface architectures and recommendations for low-cost, inclusive game engineering frameworks.
This paper introduces a reinforcement learning framework that enables controllable and diverse player behaviors without relying on human gameplay data. Existing approaches often require large-scale player trajectories, train separate models for different player types, or provide no direct mapping between interpretable behavioral parameters and the learned policy, limiting their scalability and controllability. We define player behavior in an N-dimensional continuous space and uniformly sample target behavior vectors from a region that encompasses the subset representing real human styles. During training, each agent receives both its current and target behavior vectors as input, and the reward is based on the normalized reduction in distance between them. This allows the policy to learn how actions influence behavioral statistics, enabling smooth control over attributes such as aggressiveness, mobility, and cooperativeness. A single PPO-based multi-agent policy can reproduce new or unseen play styles without retraining. Experiments conducted in a custom multi-player Unity game show that the proposed framework produces significantly greater behavioral diversity than a win-only baseline and matches specified behavior vectors across diverse targets, with highest accuracy on score-based dimensions and moderate degradation on cooperation and mobility. The method offers a scalable solution for automated playtesting, game balancing, human-like behavior simulation, and replacing disconnected players in online games.
Difficulty polarisation–where the same procedurally generated dungeon is trivial for one player and overwhelming for another–remains a persistent challenge in game design. Reinforcement learning (RL) agents can potentially adapt generator parameters to individual players, but only if those parameters demonstrably control flow-relevant output properties. This paper presents a two-scale parameterised Binary Space Partitioning (BSP) dungeon generator exposing six continuous parameters–three macro-structural (room count, lock-key depth, loop count) and three micro-pacing (safe weight, treasure weight, distance influence)–and evaluates their controllability across 12,500 generated dungeons using six flow-grounded metrics. Non-parametric sensitivity analysis (Kruskal–Wallis epsilon-squared (ϵ²), Spearman rank correlation (ρ), Jonckheere–Terpstra trend tests) reveals that safe_weight is the strongest single control (ϵ² = 0.49 for combat density gradient), lock_key_depth provides reliable macro control (ϵ² = 0.28 for progression depth fidelity), and the two scales are independently controllable by architectural construction. A synthetic-player traversal study on the same corpus provides behavioural validation: macro parameters produce large traversal effects (ϵ² up to 0.49) while micro parameters yield exactly zero traversal effect, confirming cross-scale independence under an independent evaluation. These findings demonstrate controllability over flow-grounded proxy metrics – the prerequisite for RL-based player flow optimisation – and provide a parameter leverage ranking to guide reward function design.
Heuristic evaluation is widely recognized as an effective method for assessing software interfaces. This article proposes a set of 52 heuristics specifically designed for the characterization of pervasive games, a paradigm in which entertainment goes beyond the gameplay to extend into the real world by blurring traditional game boundaries of space, time, and social interaction. Accordingly, the proposed heuristics address essential elements of a pervasive gaming experience such as interaction, immersion, and socialization, with a focus on expanding traditional game dynamics. This work is grounded in a defined set of pervasiveness properties, understood as the characteristics that enable video games to expand across spatial, social, and temporal dimensions. The heuristics draw on our experience in designing and developing pervasive games and video game evaluation methods, as well as on state-of-the-art research. They were further refined through two expert panels (comprising 11 and 12 participants, respectively) to ensure completeness and relevance. Their aim is to provide experience-based guidelines that support the design and evaluation of pervasive games, enabling the effective incorporation of the desired spatial, social, and temporal aspects of pervasiveness into new games, as well as the comparison of existing games in terms of the type, depth, and gameplay impact of their pervasiveness.
We explore long-term prediction of quitting behavior, a challenge across many domains, using League of Legends (LoL) as our test case. The objective is to present a replicable and stable methodology for long-horizon game quitting prediction when many and potentially highly collinear predictors are available, as is often the case in practice. We show that we can robustly identify long-term predictors of quitting behavior by selecting features of importance using Random Survival Forest analysis to constrain the number of Cox models subsequently used for statistical inferences. In the case of LoL, defensive deficit (a combination of class-normalized damage taken and number of deaths), computed from the early season matches (first 50), was found to consistently predict the quitting hazard over the entire season. It did so to the same extent across two different datasets, such that higher damage taken by one standard deviation, early in the season, lowered the quitting hazard by about 18% across the two datasets.
