Abstract -- Conjuring with a "Tap Tap", learner initiates building a network graph of a spatial interactive sandbox.
Start playing immediately, progressively increase the complexity for a personalized generative play experience. Augmented with Large Language Model (LLM), the infinitely adaptable sandbox enables self-directed learning paths and self-actualized objectives. It is just as important to set boundaries in this learning environment with human designed scaffolding for clear guidance and worthwhile learning goals. Remarkably, LLM can also re-organize the graph layout for alternative perspectives. These together form the enhanced augmentation of the generative learning process.
Prototype testing shows that learners found the interface intuitive and provide a substantial in-depth experience. The system even demonstrates flexibility to accommodate learning experiments in multiple knowledge domains.
A lifelong journey of learning is fueled by a genuine sense of curiosity and regular engagements in constructive thinking – both are vital for driving practical innovation. Traditional educational approaches often fall short in cultivating these traits, frequently relying on rote memorization, static context-limited tasks and standardized assessment (Ahmed 2025). This research explores how an LLM-augmented game-based system relying on human spatial intelligence can offer a more engaging, dynamic and responsive tool for learning experiments, a Learning Analytics apparatus akin to Thorndike's puzzle box, where learning processes such as Generative Learning and Self-Regulated Learning (SRL) can proceed, tested and be observed.
Generative learning model defined the brain as a model builder where understanding extends from prior experience, key being the generation of relations among concepts (Wittrock 2010). Aligning with this model, the sandbox's primary use case is the construction of a network graph by making connections at user's own discretion. In the discussion, I argue that it is possible to operate this using strictly human-computer interface components by taking full advantage of spatial thinking, followed with LLM augmentations to take care of handling the required individual non-deterministic creative process, and then back to calibrating the accuracy of the LLM with various strategies including human spatial judgments.
While SRL is fundamentally a life-long process requiring intentional practice and reflection, well-designed game can create supportive conditions for enacting and developing SRL behaviors (Zheng 2024). I argue in related work that such environment would make available most of the popular SRL strategies for both learners and educators.
Practically, this kind of visualization project is best designed iterative-ly (Munzner 2009), as described in detail in methodology. Many of the capabilities relies on open source free web technologies, which is the only way to customize and integrate everything so fully, and made this project even possible without a budget.
Throughout this writing, the terms 'user', 'player' and 'learner' are used interchangeably, the nuance difference is that 'user' is more concern with the interactive experience, 'player' is more about gamify engagement, while 'learner' is more about the learning process. Hence this project is the combination of these roles in different research areas. Finally, the paper offers reflection, observations and insights gained at the conclusion of this project.
The following research questions started the project:
Games offer an ideal medium to structure and motivate exploratory learning. Through mechanics such as goal-setting, feedback loops, and progressive challenges, games naturally engage learners in cycles of planning, action, and reflection. More specifically, word-building games like TapTap.Build implicitly teach principles of design and fabrication: user learn to combine elements, evaluate outcomes, iterate on ideas, and understand relationships within a system — much like an architect, engineer, or creative designer would. Unlike purely open-end sandbox experiences (e.g. Minecraft), educational games can embed guided scenarios that focus inquiry, integrate learning objectives, and provide scaffolds — all while preserving a sense of agency and discovery.
In this way, game-based environments do not merely entertain; they create immersive contexts where curiosity is rewarded, critical thinking is practiced, and foundational design-thinking skills are cultivated.
This dynamic aligns closely with the principles of SRL, a process by which learners take control of their own learning by setting goals, monitoring progress, and adjusting strategies as needed (Zimmerman 2008), (Zimmerman 2015). Especially for young learners, a constructivist, task-oriented learning strategy—where knowledge is built through active engagement with meaningful activities—consistently yields strong outcomes (Flavell 1979). Game environments can provide a rich context environment for SRL tasks, as they often require user to unconsciously set goals and make plans to adapt to changing circumstances in order to succeed; in addition performance can be measured with data-driven analysis of game-play (Zap & Code 2009), (Goslen 2025). When combined with learning analytics — the collection and analysis of learner data to optimize educational experiences—game-based SRL environments gain a powerful feedback loop for both learners and educators (Lai 2025).
As the prevalence playing of games online has been increasing, especially among younger population 2 shows. Web games are video games with potentially powerful medium to engage learners, web games are convenient, short bursts of interaction, easily accessible on multiple device formats. On mobile, that means anytime and anywhere. Video games have several characteristics that make them suitable for educational purposes: (1) interactive, (2) repetitive, (3) adaptive, (4) cumulative, (5) scaffold-ed, (6) affectively situated, (7) intrinsically oriented, and (8) based on both player-centered and game-based goals (Zap & code 2009).
However, designing good game-based learning requires engaging user's agency, curiosity and captivation to ensure that the game is both engaging and educationally effective (Boudreau 2021). In addition learning analytics could be used to enhance the user engagement and provide the guidance to introduce users to the system and create the learning process to mastery of using the game system.
Development of this project commenced with the re-engineering of a novel popular LLM-driven game, then improved upon by the progressively integrate pedagogical features. This iterative approach usually encompasses these working modes:
| Working Modes | Activities |
|---|---|
| Inspiration & Research | Study a popular relevant game & identify educational gaps |
| Software Integration | Chose amongst open-source tools for existing solutions |
| New Mechanics Experimentation | Implement and tested new features and interactions |
| LLM Integration | Trial and Error testing local LLM for reasonable output generation |
| Learning Analytics Layering | Add human-in-the-loop controls and monitor |
| Adhere to Design Guideline | Keeping minimalistic design for better focus |
The original game idea is inspired by Infinite Craft (Agarwal 2023), a browser game where player combines elements to craft new elements. Infinite Craft uses Llama 2 to generate elements, which allows a theoretically infinite number of combinations to discover! (Infinite Craft Wiki 2025).
The game is actively played and received positive reviews for its ability to spark curiosity (Press-Reynolds 2024). Users are attracted to the open-end-ness of the game, and the ability to discover new elements through experimentation. The game also has a simple and intuitive interface, making it easy for users to get started.
However Infinite Craft has a few limitations in regards to the research questions:
This lead to finding an alternative code base called opencraft (Bufferhead 2024) which is an open-sourced faithful recreation of Infinite Craft.
However opencraft also has limitations, but it was a good starting point to build upon:
| Limitations | Implications |
|---|---|
| System design based on assumptions about infinite craft | Inherit above same limitations |
| Vite + Vue framework for DOM based graphics | reasonable for graph with no links |
| Hard-coded to a particular LLM | Need switch-able model library to keep up with new models |
The key to guiding the development is to address the above limitations one-by-one, then using an iterative approach to refine the game via testing and feedback.
In fact, it is important to get the game out in-the-wild as soon as possible, then users give feedback to guide further development. This was done in private and also via two short class presentations to get feedback from peers.
Major code rewrite were made, LLM assistance using deepseek.ai were used to help speed up the development process especially the data collection section:
| Implemented Capabilities | Implications |
|---|---|
| 'double click' action to add any new word | Allow users to start with any word of their choice |
| New original LLM prompt | Generate real world objects instead of fantastical elements |
| '3 seconds hold' action to wikipedia page | Seamless discover word's meaning and concepts in more details |
| Rewrite back-end architecture | Newer more capable local LLM Allow combinations beyond 2 elements |
| Single 'select' action for breakdown | Engage curiosity about what constitute an element |
| Network graph visualization | Allow users to see relationships between words |
| Implement responsive design | Allow for play on different devices including mobile |
| New lesson module | Guide users towards mastery of playing the game |
| Add a data collection component | Possible to monitor user progress and engagement for learning analytics |
| Auto graph layout | Rearrange entire graph with a LLM prompt for new perspective |
| Special command line | Extend with custom executable functions |
| Game graph sharing | Allow users to share their creations with others |
The result is a mobile-responsive web application with the following major components 3:
| Component | Used | Figure |
|---|---|---|
| Front-end interface | HTML/CSS/JavaScript stack with D3.js for superior rendering performance | 1, , |
| Back-end server | Node.js to handle user requests | |
| LLM integration | Ollama.js, connect with local LLM models gemma3:4b-it-qat, model released on Sept 2025, later updated to gemma4:e4b, released April 2026 | |
| User interaction recording | RRWeb to record minute movements like mouse movement | |
| Progression checking lesson modules | Custom tutorial system written in javascript + json |
Built as an apparatus for experimentation, here we take an example to apply the Generative Learning model and discuss benefits and risks.
Generative learning model defines four main processes (Wittrock 1992): a) Attention - Directing focus to relevant incoming information, b) Motivation - The drive to engage in the generative process, c) Knowledge - Accessing and retrieving relevant prior knowledge from long-term memory, d) Generation - Constructing new relationships between new information and prior knowledge.
To "Motivate" user to virtually fabricate new concept from prior concepts, create relations between concepts and organizing the network structure; instead of a conventional text-based LLM chat that's linear and long formatted, let's go with a simple minimalistic network graph interface, creating, connecting and destroying nodes using just tap and drag gestures. Reducing the engagement effort yet still offering increasing complexity, while helping user maintain "Attention" with minimal cognitive load.
According to research, effective generative learning in practice can be implemented with Mapping (Fiorella 2015), which is spatial arrangement of words and links, familiar examples are linked concept maps, tree structured mind maps and cause-and-effect flow charts.
In our application, Mapping can be created with three simple actions: positioning, insertion and removal. With just these, analytic spatial reasoning such as encoding proximity and relational grouping would be possible; also learners can externally represent their internal mental model. This aligns with two core components of spatial thinking, namely flexibility in strategic spatial choice and meta-representation competence (Hegarty 2010).
With the memory retrieval and clarification of "Knowledge", a seamless hyperlink-like feature with tap and hold gesture looks-up a node's term in Wikipedia. This design takes homage from Ted Nelson's Xanadu Project (Nelson 1999), (Nelson 2001). Following the full principles of hypertext would be a formidable task (Voß 2019), nevertheless graph nodes backlinking brings learner to the reference source material, as a form of transclusion. Learner is expected to read the wiki article and return to the sandbox once understanding is placed in working memory.
During iterative prototyping it was found that human-designed scaffolding is an essential add-on, absolutely necessary to ensure learner exploration can be imaginative but not aimless. In fact, the best example is the application itself, user goes through a step-by-step tutorial to master above actions first before being allowed to "Generate" construct freely with LLM.
LLM provides a limitless context-rich environment by generating new nodes on demand. With tap selections of two or more nodes, a new combination prompt is passed to the LLM, which dynamically response with a new set of nodes with connections of the most probable outputs from combining the selected ingredients. Existing nodes represents learner's prior pinned "Knowledge", and this augments the "Generation" process. Similarly, tap a single node decomposes into new set of component nodes, showing the required ingredients.
It would be easier to actually try these mechanisms than to explain in words, the key to the Augmenting Generative Learning chest box — one that goes far beyond predefined context and always catching up to learner's edge of curiosity, augmenting "Motivation" with personalized scaffold-ing at the Zone of Proximal Development (Cai 2024), literally it is real time overlapping and expanding of learning boundaries.
Moreover augmentation of spatial thinking is possible with LLM suggesting a new layout. When user issues command :remix with a hunch, the LLM takes the entire graph and gives back a possible new layout; a new spatial structure with a new perspective. This is meta-cognitive activity, not of the knowledge itself, but the organization of. Below is an example for language learning, is an example in science inquiry.
►
►
More complex examples might be where learner is interested in the proportions of the ingredients involved, closer nodes have higher proportions, further nodes contributes less. Or perhaps the positions of resulting nodes depends on the difficulty of fabrication or energy usage as below .
▼
While LLM will always carries the risk of hallucination (Adam 2024), corrective calibrations can be put in place. Numerous strategies exist (Huang 2025), (Islam 2024), but lets focus on ones which taptap.build has directly used.
A straight forward strategy is to carefully design the node generation prompt - more or less a time consuming trial and error process. Nevertheless we ended up with below well tested prompt structure as a strong first line of defense:
| Prompt Section | Primary Function | Rationale |
|---|---|---|
| Contextual Anchoring | Establishes domain-specific persona and task framing | Sets baseline of what it needs to do |
| Semantic Boundary Setting | List permissible and non-permissible scope of reasoning | Clarifies the kind of meanings we desire and despise |
| Formatting Constraint | Enforces a strict list of single items | Requires the model to response only in essential atomic units |
| Avoid Common Hallucinations | Lists specific, observed failure modes in point-form | Explicit negative exemplars is much better than generic "don't hallucinate" |
| Reaffirm Negatives | Reinstate non-permissible output terms | Readdress semantic boundary with more actual observed wrong outputs |
| Reaffirm Formats | Re-iterates format constraints in a different syntactic form | Remind LLM to not veer off to non-relevant formats |
| Simple Examples | Provides a list of desired input-output behavior | Establish concrete demonstrations to follow |
Contributing the most to stable accuracy is the formatting constraint, which enforce output to only the essential words, effectively limiting the rhetorical abilities of LLM; this means no way for LLM to introduce convincing wordings to mislead user and limiting it to give focused answers. Uniquely, this only works for spatial graph which gives a near instant comprehensive overview of relations without using additional words, avoiding wordy conversational chat.
Reaffirming conditions in different ways also seems to help, even if its obvious to human by first definition. LLM is still a blackbox, perhaps it has something to do with the model indecisiveness with conflicting parameters, or perhaps being creative means trying to be flexible with rules, no matter the reason, it means that forcing it to consistently "act" a specific way is somewhat against its design, deliberate calibration of all edge cases is required and it is unlikely a single prompt can ever achieve this. Nevertheless here is an attempt.
Generate complexity but answer simply.
To know when it is "right", it has to know when it was "wrong".
"No" is considered, but a "No ... Don't ... " definitely means no.
The next calibration are the piecemeal information that feeds the network, generated from mutually isolated contexts makes it hard for LLM to compound and propagate lies. Even if an individual part is wrong, user will be going broader and further down in detail, especially when manually fact-checking using external evidence - a built in feature. Although demanding tedious human cognitive effort, our spatial intelligence makes it much easier to spot out-of-place words and isolate inconsistent elements. Thus users check the graph through deductive reasoning, this recursive decomposition of claims is similar to academic reference checking or transclusion, eventually exposes any lies.
The sum of the whole is more than the sum of the parts.
Turtles all the way down is vicious.
Further calibration becomes increasingly abstract. The system addresses this through a :whatif prompt — a structured narrative approach that :remix the entire graph. Asking the LLM if the graph could have a different structure allows the user to seek a different perspective, switching mental model that potential gives better answers. Because The graph is already casually linked - a spatial chain-of-thought, the graph already forms the basis for casual explanation (Amir 2021), by rearranging the nodes spatially with guidance, we can directly test counterfactual possibilities and find different explanations, :whatif is a prompt that works as follows:
| Prompt Section | Primary Function | Rationale |
|---|---|---|
| Narration Anchoring | Establishes graphic designer persona and narrative framing | Sets baseline of what it needs to do |
| Typical Story Rules | List what others commonly expects to find in an infographic | Clarifies audience expectations and conventional patterns |
| Interpretation | Listen to what the story currently tells | Try to really understand what is already going on |
| Mood Reading | Assesses the overall the shape and tone of the story | Help understand if we can do it differently |
| Main Point Prioritization | Rank and prioritizes core ideas | Determines what matters most and what should take precedence |
| Visual Style Selection | Determine what style approach should be used | How to tell the new story |
| Apply as Coordinates | Consolidate above and execute | Coordinates per item is the end goal |
| Explain | Include a brief explanation of reasoning | Give reason for behavior |
| Formatting | Provide current structure to get output structure | Give this last for closest context |
Here, the structure of the prompt is much more about presentation from existing elements, hence the story telling approach was taken.
Read the current story and tell a different one.
The impact of above strategies increases as the network grows, because of more and more conceptual guidance with human-in-the-loop (Wu 2022). By first giving correct words, then securing various parts of the graph, there is now a foundation to further verify the top-level claims, going higher in abstractions and the network becomes more trustworthy as it grows while remaining comprehensible such that conceptual exploration can thrive responsibly.
This report presents the development journey of TapTap.Build, what started as a game ends up as a learning apparatus supported by Ai. In this, learners combine words through tapping and dragging to intuitively manipulate a visual concept map. Ai assist in generating new word combinations and network structures in real time, acting as a knowledgeable partner in the exploration of an infinitely vast learning environment.
Being web-based makes this immediately accessible on all screens and available at all times, which might seem trivial, but it means free pursue of innovation and learning at convenience, especially suitable for classrooms tryouts without the need for long complicated setup. I lost count the number of times I linked concepts virtually on the go, testing new combinations without formulating full prompt is liberating, and scratches the itch to construct or break things down virtually right away. There is certainly feeling of freedom of expression in the moment.
That is why this should appeal to all ages, especially for learners whom still have highly creative minds, requiring only rudimentary command of digital device use. Even vocabulary or grammatical skills are not needed, in fact these can get developed as per discussion. From this, every interaction helps user grasp new concepts via personally motivated conceptual world building.
Although LLM is providing the open-end-ness to allow rich context. This boundlessness inevitably brings about confusion, distraction and misdirection.
Very early on, it was clear from user feedback that they were confused and did not even know where to start. A tutorial lesson module was developed just to introduce the system, step-by-step, and this human designed scaffolding was found to be effective.
As for distraction, selective attention becomes paramount and this is where a graph is better than the typical Ai chat interface. With the verbose chat interface, information overloads our limited working memory (Miller 1956). Instead let's rely on the innate human spatial reasoning to handle the complexity, with an intentionally minimal graph design relative to say a 3D simulator, the goal is to maximize attention and minimize distraction, see Principles of Data-Ink Ratio (Tufte 1983), where good graphical representations maximize data-pixels and erase as much non-data-pixels as possible.
As for misdirection, Ai can be spectacularly misleading, we human must rein in Ai hallucination in order to ensure information is correct, because we will be responsible for any real outcome. I am sure there are more ways than the ones detailed in the discussion, because everyone has dealt with some forms of deception before, be vigilant and use our strengths, in this case spatial intelligence and questioning. It is our provocative, in fact itself a fun learning process, to poke holes in Ai's confident bubble and correct it for our own benefit.
Looking back, these mitigations are absolutely essential, they came about from necessities. Central to the human intervention are teachers, by their close physical attendance and instructional duty to students, they are in the best position to blend this with other non-virtual effective generative learning activities (Fiorella 2016), for example supporting students to teach each others about their own graph as a learning strategy similar to PeerSynergy (Sam 2024). Actually, teachers could see more risks and experiment with tricks on the front line of this cat and mouse game.
The methodology used is shown both as a study of how to rapidly develop a prototype apparatus and the design reasoning taken. As the journey is the work of a one-person journey typical of solo-entrepreneur (Basuthakur 2024). Having prior industry experience of having developed software projects (Lee 2025) provided the necessary technical skill-sets needed to push progress towards realizing a working prototype. Reflecting on the advantages, disadvantage of this development process:
| Advantages | Disadvantages | Mitigations |
|---|---|---|
| Quick decision making | Limited perspective | Peer feedback |
| Deep ownership | Limited resources | Modularization and rely on open source solutions |
| Flexible project direction | Goal deviation | Documentation and review |
| Rapid prototyping | Risk of burnout | LLM coding assistance |
As this is a prototype study, we will leave more rigorious experimentation to further research, especially in learning analytics. This project is in a sense a precursor of such experimentation, building the necessary tool-chain. I expect future experimentation to be based on different models and knowledge domains, turns out the same tutorial system can also be customized by researchers to add instructions, conditions, constraints and assessments for other learning experiments.
We can already see that this system is definitely tailored to science and technology innovation, but it is also suitable for language learning, starting with vocabulary building. As users can add new words and learn their meanings via the seamless learning and nuances with spatial relationships. Research has shown that games to be particularly useful for English as a Foreign Language (EFL) learners (Soyoof, Reynolds, Chan, Tsend & McLay 2025).
Ultimately, TapTap.Build serves as a fun and engaging testbed for how Ai can stimulate more inquisitive, challenging and self-directed approach to learning. It has developed to a point where it contains the necessary features to further explore more deliberate learning interventions and to analyze their impact.