The motivation behind the project is a persistent gap in the artificial intelligence landscape. Although open-source large language models have advanced rapidly in recent years, most of the research community’s effort has poured into general-purpose models trained predominantly on English data. That imbalance creates real problems for anyone studying or deploying AI in Chinese education. Chinese is a language with distinctive tokenization challenges, rich morphological structure, and a vast educational literature that English-centric models simply do not capture well. Educational applications add another layer of specificity: the vocabulary of pedagogy, curriculum standards, exam questions, and subject-specific reasoning in Chinese differs substantially from the web text that dominates most training corpora.

Rather than fine-tuning an existing model, the team took the more demanding route of training CELLM from the ground up, a decision that gave them full control over the data pipeline and the architecture. The process unfolded in two stages. The first was pre-training, in which the model learned the statistical structure of language from an open-source dataset drawn from the Chinese education domain. During this phase, the model absorbed the patterns of educational Chinese text, building the foundational knowledge that later stages would refine. The second stage was instruction fine-tuning, which teaches a base model to follow directions, answer questions, and behave like an assistant rather than a text predictor.

For that second stage, the researchers faced a familiar obstacle: high-quality Chinese instruction data for education was scarce. Their response was to build their own. They constructed a Chinese instruction dataset comprising over 258,000 data entries and, in the spirit of the entire project, released it openly. Instruction datasets of this kind typically pair prompts with high-quality responses, spanning formats such as question answering, explanations, and multi-turn dialogue. By curating the data themselves, the team could steer the model toward the kinds of interactions that matter in educational contexts, from explaining a mathematics problem to discussing language teaching strategies.

The architecture underlying CELLM draws on the core technologies that have defined modern open-source language models. The team reviewed and synthesized the design choices of representative open-source systems before settling on their own configuration. Among the techniques reflected in the model’s lineage are rotary position embeddings, an approach that encodes word order by rotating vector representations and has become a staple of contemporary transformer designs. The model also builds on advances in attention mechanisms, including grouped query attention, which reduces computational cost by sharing key-value projections across multiple query heads, a strategy popularized by efficient inference research.

Other components of the design reflect a decade of accumulated transformer engineering. The feed-forward layers employ variants of gated linear units, activation functions shown to improve transformer performance over standard alternatives. Training at scale was supported by DeepSpeed, the distributed training framework developed to make models with hundreds of billions of parameters feasible on real hardware. The team also drew on research into scaling laws and overtraining, which examines how model performance grows with parameters and data, informing how a relatively small model can be trained to punch above its weight class.

The choice of a 1.5-billion-parameter size is itself significant. In an era when frontier models boast hundreds of billions of parameters, a compact model might seem modest. But smaller models are dramatically cheaper to train, run, and deploy, which matters enormously for educational research groups, schools, and developers working with limited computing budgets. A 1.5-billion-parameter model can run on consumer-grade hardware, making it practical for experimentation and classroom applications alike. The trade-off is capability, and the evaluation results provide a transparent accounting of where the model stands.

That transparency is one of the study’s most valuable contributions. The researchers benchmarked CELLM across multiple evaluation datasets, including suites designed to measure massive multitask language understanding in Chinese, such as C-Eval and CMMLU, which test knowledge across disciplines and difficulty levels, alongside mathematical problem-solving benchmarks. The published results establish a reference baseline for future research, giving subsequent teams a clear point of comparison. In a field where claims often outpace evidence, a fully documented baseline for an education-specific Chinese model fills a genuine need.

The open-source release strategy amplifies the work’s potential impact. Everything generated during the study, including the models, the data, and the code, has been made publicly available. This stands in deliberate contrast to the closed ecosystems of the largest commercial AI systems, whose training data and internal workings remain opaque. Open release allows other researchers to scrutinize the model, reproduce the results, extend the training, or adapt the instruction dataset to adjacent domains. It also enables studies of the models themselves, including investigations of bias, generalization, and how a model’s capabilities trace back to its pre-training data, questions that are difficult or impossible to answer with proprietary systems.

For the field of Chinese education research, CELLM represents both a tool and a template. As a tool, it offers researchers a capable, domain-tuned language model they can study, fine-tune, and deploy without licensing barriers. As a template, it demonstrates that building a specialized model from scratch, with a purpose-built instruction dataset and rigorous public evaluation, is achievable by a small academic team. As open-source language models continue to proliferate across languages and domains, the Shanghai team’s work signals a shift toward AI research that treats education not as an afterthought of general-purpose systems, but as a domain deserving models, data, and benchmarks of its own.

Subject of Research: An open-source large language model trained from scratch for Chinese education research

Article Title: An Open-Source Large Language Model for Chinese Education Research

Article References: Liu, W., Hao, H., & Zhou, A. (2025). An Open-Source Large Language Model for Chinese Education Research. Frontiers of Digital Education, 2(2), Article 23. https://doi.org/10.1007/s44366-025-0060-0

Image Credits: AI Generated

DOI: 10.1007/s44366-025-0060-0

Keywords: large language models, open source, Chinese education, CELLM, instruction fine-tuning, pre-training, transformer architecture, rotary position embeddings, DeepSpeed, C-Eval, CMMLU, educational AI

The challenge the team set out to solve is one that any designer of intelligent tutoring or e-learning systems will recognize. Human learning behavior does not present itself in clean, uniform data streams. A learner may click rapidly through a page, then pause in silence for several seconds, issue a brief voice query, or frown at a difficult passage. Each of these channels—clicks, speech, gestures, and micro-expressions—carries different information, and the reliability of each channel changes over time. Speech may drop out, a camera may partially lose track of a face, or clicking may go quiet during deep concentration. Traditional fusion pipelines, which assign fixed weights to each input type, tend to stumble when this balance shifts, and static attention mechanisms struggle to respond to the dynamic texture of real learning sessions.

MFT-Net attacks the problem with three interlocking components built on a Transformer encoder backbone. The first is a modal response filtering module that sits before the main encoder. Rather than tokenizing the entire behavioral stream uniformly, it computes an overall multimodal response intensity at each time step and applies a threshold, tuned on validation data, to separate salient behavioral moments from low-activity noise. This means short but meaningful events—rapid backtracking, a brief hesitation, a quick gesture tap—are preserved instead of being diluted by thousands of uninformative frames. When a modality goes dark for several consecutive steps, the system fills the gap through behavioral time alignment and modal interpolation, keeping the input tensor structurally consistent.

The second component is the one that gives the framework its name: a dynamic modality weight regulation network. Each modality is first projected into a shared latent space, and cross-modal consistency is then estimated by measuring pairwise Euclidean distances between modality embeddings. Modalities that agree with their neighbors—suggesting they are picking up the same behavioral signal—receive larger contribution weights, while noisy or poorly aligned channels are suppressed. The weights are normalized so that their sum equals one, preventing any single channel from dominating the fused representation. The researchers stress that consistency is treated as a proxy for reliability, not a direct measure of it, and they compared Euclidean distance against cosine similarity and a learned attention-based metric, finding that the Euclidean approach offered the best balance between recognition performance and inference efficiency.

The third innovation concerns how the model understands its own output categories. Learning behaviors are organized in a hierarchical dictionary of 18 labels: six coarse families such as navigation and interaction, information seeking, affective response, collaboration, off-task behavior, and hesitation, plus twelve fine-grained atomic labels including click-scroll, voice-query, gesture-tap, long-dwell, repeated-backtrack, and silence. Instead of treating these labels as arbitrary identifiers, MFT-Net converts them into structured 64-dimensional embedding vectors and injects them into the attention-based matching between behavioral sequences and categories. A label-guided semantic projection uses the label embeddings as queries against the sequence representations, allowing the model to highlight the parts of a behavioral stream that align semantically with each candidate label. This helps separate semantically close categories that would otherwise be confused—for example, distinguishing genuine hesitation from simple silence.

The mathematical machinery beneath these modules follows the familiar Transformer recipe. Multimodal features are fused as a weighted sum of per-modality embeddings, position encodings are added to preserve temporal order, and multi-head attention extracts global dependencies across the behavioral sequence using the standard scaled dot-product formulation. The classification task is cast as a single 18-class softmax problem with hierarchical decoding, so invalid parent-child combinations cannot occur, and cross-entropy loss is applied over the structured label space. Deployment considerations shaped the design as well: attention-channel pruning and low-rank compression of the label embedding dimensions reduced the serialized model to 18.6 megabytes on a workstation GPU and 9.4 megabytes in a compressed edge configuration.

Experiments were conducted on a dataset of 5240 sequence-level multimodal learning-behavior samples drawn from 312 learning sessions, with click traces, speech cues, gesture records, and facial-expression features synchronized at 30 frames per second. Crucially, the data were split at the session level using stratified group splitting, so sequences from the same learning session never appeared in both training and test sets—a safeguard against leakage that many behavior-recognition studies overlook. Training used AdamW with a cosine learning-rate schedule, and the full model converged to its peak accuracy in just 8 epochs. Against five baselines—a shallow MLP, Bi-GRU with attention, a static-fusion Transformer, Transformer-XL, and ConvLSTM—MFT-Net achieved 96.2 percent accuracy and a 95.5 percent F1 score, along with an average inference latency of 34 milliseconds.

Robustness testing revealed perhaps the most practically important results. Under a skewed test distribution with deliberate label and modality imbalance, MFT-Net maintained 90.6 percent accuracy at 33 milliseconds per inference, while Bi-GRU with attention fell to 84.3 percent with latency rising to 59 milliseconds. In leave-one-modality-out tests, removing the click channel caused a larger performance drop than removing the gesture channel, indicating that interaction traces carry especially strong behavioral evidence. The model also outperformed all baselines in single-modal and dual-modal settings, achieving 84.1 percent accuracy with one channel and 90.2 percent with two. Repeated runs with paired statistical tests confirmed that the improvements over every baseline were significant, with Holm-Bonferroni corrected p-values below 0.001 and large paired effect sizes.

The authors are candid about the framework’s limitations. Modal response filtering depends on threshold and window settings that can suppress informative events if too strict or admit noise if too loose. The Euclidean similarity measure can be biased by embedding scale if normalization is inadequate, performance may degrade when multiple modalities fail simultaneously, and cross-setting transfer—training on desktop logs and testing on mobile tap streams—still produced measurable degradation. Future work, they write, will focus on more adaptive thresholding, uncertainty-aware similarity metrics, and hardware-specific validation for mobile and edge deployment.

Even with those caveats, the study offers a compelling blueprint for the next generation of adaptive learning platforms. By coupling temporal salience filtering, consistency-driven modality regulation, and label-aware semantic matching in a single end-to-end pipeline, MFT-Net demonstrates that AI systems can read the messy, shifting, multimodal texture of human learning behavior accurately enough—and fast enough—to provide meaningful personalized feedback in real time. For digital education, where a missed moment of hesitation or an unnoticed gesture can mean the difference between timely help and a struggling learner, that capability may prove transformative.

Subject of Research: Multimodal deep learning for adaptive learning behavior recognition using dynamic modality regulation in a Transformer architecture.

Article Title: Adaptive learning behavior recognition using multimodal transformer based dynamic modality regulation

Article References: Zujun, L., Xueyao, D., & Faze, L. (2026). Adaptive learning behavior recognition using multimodal transformer based dynamic modality regulation. Discover Artificial Intelligence, 6(1), Article 1184. https://doi.org/10.1007/s44163-026-02112-3

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02112-3

Keywords: multimodal transformer, learning behavior recognition, dynamic modality regulation, structured label embedding, adaptive fusion, attention mechanism, multimodal fusion, behavioral sequence modeling, intelligent learning systems, MFT-Net, human-computer interaction, educational AI

Knowledge tracing is a family of computational methods designed to estimate a learner’s evolving mastery over time. In a typical system, each exercise is linked to one or more knowledge components, such as solving a linear equation, applying a grammatical rule or identifying a chemical property. The model receives a stream of responses and updates an internal estimate of the learner’s knowledge after each interaction. Traditional Bayesian knowledge tracing represents learning as transitions between states, often balancing the probability that a student has learned a skill against the possibility of guessing, making a careless mistake or forgetting. More recent approaches use neural networks and other machine-learning architectures to capture complex patterns in long sequences of student activity. These systems have generally emphasized temporal information: what happened previously, how much time has passed and how a learner’s performance changes across attempts. The study’s authors note that this focus leaves another source of information comparatively underused—the spatial organization of the knowledge itself. Here, spatial does not mean physical distance. It describes the relational arrangement of concepts in a knowledge graph, where edges indicate dependencies, influence or causal connections among components.

The distinction matters because educational knowledge is often hierarchical. Understanding how to solve a two-step equation may depend on addition, subtraction, multiplication, division and the ability to preserve equality across operations. In a science course, interpreting a graph may require knowledge of axes, variables and proportional relationships before a student can reason about a more advanced model. If an assessment system examines only the label attached to the current exercise, it may overlook the chain of concepts that supports performance. A network-based system can represent these dependencies explicitly. In such a representation, a knowledge component is a node, while a directed connection can indicate that changes in one component are related to another. Immediate neighbors form a first-order structure. A second-order structure includes connections reached through those neighbors, allowing the model to consider a wider local context. The researchers studied whether these multilevel relationships could improve knowledge tracing models across both deep-learning and traditional machine-learning settings. Their analysis treats the structure as informative evidence rather than as a decorative visualization added after prediction.

To construct the relevant relationships, the researchers used causal structure learning, a group of statistical methods that attempts to infer directional connections from observed variables and their patterns of association. In educational data, the variables can represent knowledge components and the response information associated with them. Causal discovery does not automatically prove that one concept directly causes another in the psychological sense, but it can provide a principled way to estimate a network of dependencies from data. The resulting structure was then incorporated into knowledge-tracing models. This step is technically important because it changes the information available during prediction. Instead of relying solely on a student’s past response sequence, a model can also aggregate signals from related concepts and from concepts linked through an additional level of the network. A learner’s performance on one skill can therefore be interpreted alongside evidence about prerequisite or neighboring skills. The approach gives the model a way to distinguish a narrowly isolated weakness from a broader pattern that may arise when several connected components remain uncertain.

The reported experiments found that second-order spatial information consistently improved model performance. The source article does not present the result as a replacement for temporal modeling; rather, it shows that spatial structure can complement the time-based information already central to knowledge tracing. That combination is significant because learning is both sequential and relational. A student’s latest answer depends on prior practice, but it can also reflect the status of concepts connected to the current task. A model that captures only one of these dimensions may miss part of the explanation. The researchers tested the structural information in deep-learning models as well as traditional machine-learning models, indicating that the benefit was not confined to a single algorithmic family. The finding suggests that educational prediction systems may gain from improved representations of knowledge even when their core predictive machinery differs. It also reframes model development: progress may depend not only on building larger or more complicated networks, but on supplying models with a more meaningful description of the domain they are trying to understand.

Performance, however, is only one part of the study’s contribution. The researchers also examined interpretable features to clarify how spatial information shaped diagnostic predictions. Interpretability methods are intended to show which inputs most strongly influence a model’s output, helping researchers and educators investigate why a system reaches a particular conclusion. In this context, a spatial feature might represent information propagated from a related knowledge component or from a concept several links away in the inferred structure. If such a feature contributes strongly to a prediction of difficulty, it may indicate that the current error is connected to a weakness elsewhere in the knowledge network. This is different from simply reporting that a student answered an item incorrectly. It offers a possible explanation for the prediction and can make automated recommendations easier to scrutinize. The article identifies interpretable analysis as a route toward understanding the factors underlying students’ learning challenges, while the associated keywords identify explainable artificial intelligence and SHAP, a method commonly used to examine feature contributions. The practical value lies in connecting prediction with a diagnostic account.

That account could support more targeted instruction, although the study does not claim to have demonstrated outcomes in classrooms. A tutoring system informed by hierarchical dependencies might recommend reviewing a prerequisite rather than assigning more exercises that repeat the same surface-level task. It could also avoid treating every wrong answer as an independent event. Suppose a student repeatedly fails problems involving a particular advanced operation while also showing uncertainty on its prerequisites. A spatially aware model could flag the connected pattern and help an instructor decide whether to revisit foundational material, change the explanation or provide practice that bridges the concepts. The system might likewise identify cases in which a student has mastered supporting skills but is struggling with a specific application. Those distinctions matter for adaptive learning because effective feedback depends on the source of an error, not merely its existence. The researchers describe the spatial perspective as having potential to inform more effective instructional strategies, but the appropriate use of such predictions would still require validation with teachers, learners and real educational interventions.

The work also leaves important questions for future research. Inferred relationships can reflect the quality and scope of the data used to learn them, and knowledge dependencies may differ across curricula, age groups, subjects and populations. A hierarchy that describes one mathematics course may not transfer directly to another, while relationships in language learning or programming may be less strictly hierarchical. Student behavior can also be influenced by factors that a knowledge graph does not capture, including motivation, fatigue, unfamiliar wording and access to prior instruction. Better structural information should therefore be treated as one component of a broader assessment system, not as a complete representation of learning. The study provides evidence that second-order relationships can improve knowledge-tracing performance and make predictions more interpretable. Its larger message is that educational AI should model the architecture of knowledge as carefully as it models the passage of time. By combining learner histories with causal and hierarchical maps of concepts, future systems may move closer to diagnosing how understanding develops—and where the next useful lesson should begin.

The result is especially relevant to the distinction between prediction and diagnosis in educational data mining. A model can become better at forecasting a response without revealing whether its estimate reflects durable learning, temporary performance, or an unresolved dependency elsewhere in the curriculum. By examining contributions from spatial features, the study provides a way to investigate whether a prediction is being driven by the assessed component itself or by information carried through connected components. This makes the inferred structure potentially useful not only as an input to a predictor, but also as an object for examining the model’s reasoning.

At the same time, the study’s causal terminology requires careful interpretation. Causal structure learning produces an inferred pattern of directional dependencies from data; it does not by itself establish that mastering one knowledge component will produce mastery of another. For instructional use, those inferred links would therefore be most defensible as hypotheses about relationships to test through assessment and intervention. The strongest near-term application is likely to be prioritizing which connected skills deserve further examination, rather than automatically prescribing a specific remedy. This distinction can help prevent a system from converting a statistical association into an unwarranted claim about the source of a learner’s difficulty. It also creates a foundation for future studies comparing structurally informed predictions with teachers’ diagnoses and students’ subsequent learning outcomes.

Subject of Research: Hierarchical knowledge structures in knowledge tracing

Article Title: Unveiling the Impact of Hierarchical Knowledge Dependencies on Knowledge Tracing: A Spatial Structure Perspective

Article References: Wei, Y., Jia, R., Ding, Y., & Jiang, B. (2026). Unveiling the Impact of Hierarchical Knowledge Dependencies on Knowledge Tracing: A Spatial Structure Perspective. Frontiers of Digital Education, 3(3), Article 23. https://doi.org/10.1007/s44366-026-0097-8

Image Credits: AI Generated

DOI: 10.1007/s44366-026-0097-8

Keywords: knowledge tracing, educational AI, learning analytics, knowledge graphs, causal structure learning, explainable AI, adaptive learning, student modeling, Unveiling, Impact, Hierarchical, Knowledge