Procedural content generation has the potential to increase a game’s replayability, reduce development costs and time, and tailor experiences to users. The generation of dialogues and narratives has gained increased interest thanks to the advance of Large Language Models (LLMs) and their potential to generate convincing texts about many subjects. However, many reported implementations have poor reception from users and face limitations. This study presents a systematic mapping of the literature on the use of LLMs for generating dialogues and narratives in digital games, aiming to discover the most used models, datasets, prompt elements, fine-tuning 1 strategies, integration methods in games, how each study was evaluated, and their main issues. Through an extensive search across multiple databases and the application of rigorous inclusion and exclusion criteria, we analyzed 55 articles that apply LLMs in gaming contexts. The most recurrent challenges include narrative incoherence, repetitiveness, memory limitations, and latency, all of which directly impact player immersion. The results reveal the predominance of Generative Pre-trained Transformer (GPT) family models and an increasing use of LLMs in game modifications and interactive environments. This work contributes to the field by providing a comprehensive overview of existing approaches, identifying research gaps, and suggesting future directions, such as the need for standardized evaluation methods, the development of robust solutions to address memory issues, and the importance of comparing LLM-generated content with human-created material.

The low recycling rate in emerging countries contrasts with the efforts of public policies and educational initiatives, highlighting the need for new approaches to environmental awareness. This study proposes a serious game called “Sustainable Planet: Mission Preserve” as an educational tool for teaching sustainable practices. The game emphasizes solid waste management and proper disposal, engaging players in challenges that simulate these actions. To analyze the game's awareness potential, we conducted a quasi-experimental one-group study (N = 32) using pre- and post-interaction questionnaires with the game. This allowed us to assess its impact on participants' ability to classify waste type and its correct disposal. The results indicate a significant increase in participants' awareness, contributing to educational technologies for environmental education.

One of the primary objectives of gamified education is to promote students' behavior change to provide a positive learning experience. Most studies, however, focus exclusively on studying how students experience the systems, ignoring the emotions generated by sensory information. To fill this lack, we conducted a between-subjects controlled experiment (N = 62) to analyze the effects of gamification on students' basic emotions (i.e., happiness, sadness, anger, surprise, fear, and disgust). Based on descriptive and inferential statistical techniques, we compared the students' basic emotions between participants who used a gamified version of a learning management system (experimental group) and a group that used the system with identical educational tasks without gamification (control group). The primary results reveal that anger was the most frequently observed emotion in both system versions, whereas scare was the least expressed emotion in both versions. No significant differences in emotions between groups were registered. We contribute especially to the fields of educational technologies and gamification with insights into the effects of gamification on students' basic emotions.

Beyond 400 million people in the world have moderate or complete hearing loss and typically communicate with sign language. Learning sign language can be challenging due to the lack of easily accessible learning materials and only a few people know their country's sign language. We present Roguelibras, a Roguelike Deckbuilder serious game that uses the genre's mechanics and gameplay as a mean to learn sign language and can be played even on low-end computers, aiding its use in developing countries. We tested the game with 29 hearing participants, using pre- and post-test questionnaires asking about the meaning of 9 different signs. According to a 5-point Likert scale question, the majority stated they enjoyed the game, while their score significantly improved when comparing knowledge of the signs before and after playing. This shows the potential of the game as an entertaining alternative to help learn sign language.

We propose a system able to orchestrate the procedural generation of three creative facets (levels, narrative, and rules), using quality-diversity algorithms to promote diversity while converging to the desired solutions and able to satisfy different player's profiles. We aim to facilitate the project's reproducibility and extension by presenting the system's architecture and high-level design decisions and showcasing how we decoupled it enough to change algorithms. Our generation starts with a grammar-based quest generator that works as a frame for two quality-diversity algorithms: a level generator, which also creates locked-door puzzles, and an enemy generator, which selects their visuals based on their weapons. A rule-based profile analysis uses a pretest questionnaire to classify the player into four profiles, and the quest generator creates different content for each of them. We use a game prototype to test our system with over 83 players, playing over 204 levels and answering a posttest questionnaire after each. Results show the game and contents were well received, and there was a difference in perception between profiles. Players who played content based on their profile enjoyed it more than those playing content from another profile.

Action-Adventure games have several challenges to overcome, where the most common are enemies. The enemies’ goal is to hinder the players’ progression by taking life points, and the way they hinder this progress is distinct for different kinds of enemies. In this context, this paper introduces an extended version of an evolutionary approach for procedurally generating enemies that target the enemy’s difficulty as the goal. Our approach advances the enemy generation research by incorporating a MAP-Elites population to generate diverse enemies without losing quality. The computational experiment showed the method converged most enemies in the MAP-Elites in less than a second for most cases. Besides, we experimented with players who played an Action-Adventure game prototype with enemies we generated. This experiment showed that the players enjoyed most levels they played, and we successfully created enemies perceived as easy, medium, or hard to face.

• Single Evolutionary Algorithm encoding able to create procedural dungeon and mission. • Accurately create feasible dungeons of many sizes, linearity and locked-door puzzles. • A game prototype was developed and used for experiments with players. • Users evaluated the created content as human-made and fun as classic game levels. • Generated content allows a more complete exploration when compared to classic levels. The present research introduces an evolutionary algorithm able to procedurally generate dungeon maps containing locked door missions. The evolutionary algorithm evolves a tree structure, which encodes dungeons, aiming to generate levels close to the input configuration provided by a game designer. The tree structure holds information about the number of rooms, connections between them, and their position within a 2D map. The proposed encoding also allows evolving semantic information about the narrative of the game. This is done by feasibly setting keys and locks throughout the dungeons for locked door missions. The generated dungeons are then evaluated computationally and as a proof of concept using an adventure game prototype. A total of 70 players evaluated the contents and the results show that the procedurally generated levels are perceived as more human-made, fun, and difficult than their human-made counterparts for most cases.

The present paper proposes an evolutionary algorithm for procedural content generation of dungeon maps together with locked door missions. The algorithm evolves a tree structure which contains information of a dungeon. The aim is to converge the generated dungeons as close as possible to the input configuration set by a game designer. The dungeon holds information about rooms such as their number, connections between them and position in a 2D map (also knows as grid). It also contains relevant semantic information for generating narrative properties in the dungeon. Those are the placement of keys and locks in it, in a feasible way. Results show the algorithm is able to create dungeons within the desired configurations for a large set of different inputs. Also, they show the generated maps are perceived as human-designed, and evoke similar opinions of fun and difficulty when compared to human-designed maps.

Procedural content generation (PCG) systems are designed to automatically generate content for video games. PCG for physics-based puzzles requires one to simulate the game to ensure feasibility and stability of the objects composing the puzzle. The major drawback of this simulation-based approach is the overall running time of the PCG process, as the simulations can be computationally expensive. This paper introduces a method that uses machine learning to reduce the number of simulations performed by an evolutionary approach while generating levels of Angry Birds, a physics-based puzzle game. Our method uses classifiers to verify the stability and feasibility of the levels considered during search. The fitness function is computed only for levels that are classified as stable and feasible. An approximation of the fitness that does not require simulations is used for levels that are deemed as unstable or unfeasible by the classifiers. Our experiments show that naively approximating the fitness values can lead to poor solutions. We then introduce an approach in which the fitness values are approximated with the average fitness value of the levels' parents added to a penalty value. This approximation scheme allows the search procedure to find good-quality solutions much more quickly than a competing approach-we reduce from 43 to 25 minutes the running time required to generate one level of Angry Birds.