UniEDU: Toward Unified and Efficient Large Multimodal Models for Educational Tasks

Zhendong Chu, Jian Xie, Shen Wang, Zichao Wang, Qingsong WenEMNLP 2025

TL;DR

This paper introduces UniEDU, a unified and efficient large multimodal model designed for a suite of K-12 educational tasks, including knowledge recommendation and tracing. UniEDU significantly reduces computational overhead (3x efficiency gain) by compressing interaction histories, achieving competitive performance with minimal degradation compared to fully fine-tuned models.

What Problem Does This Paper Address?

Current AI in education often relies on task-specific models that struggle to generalize across diverse educational scenarios. Real-world K-12 materials are inherently multimodal (text and images), and existing models often focus on plain text. Furthermore, while large multimodal models (LMMs) show promise, their high computational cost, especially with long interaction histories, hinders practical deployment. This paper addresses the need for a single, efficient, and adaptable multimodal model that can handle various educational tasks and large volumes of student interaction data.

Methods

UniEDU is built upon two core modules: a Profile Encoder and a Language Model.

  • Profile Encoder: This module compresses a student's interaction history (questions, answers, knowledge concepts, correctness, time) into a compact set of tokens. This compression is crucial for managing the computational burden of long histories.
  • Language Model: This module takes the compressed history embeddings and task instructions to generate task-specific outputs. UniEDU reformulates diverse educational tasks into a generative framework, allowing for a unified approach.

A key innovation is the compression of interaction histories. Instead of feeding the entire history, UniEDU represents each interaction with a small number of compressed tokens. This significantly reduces the sequence length fed into the language model, leading to substantial VRAM savings during both training and inference. The model is trained using a standard auto-regressive objective for multi-task learning.

Datasets

The paper uses a dataset collected from a real-world e-learning platform, focusing primarily on mathematics. The dataset includes statistics on students, knowledge concepts, questions, and interactions. For training, 24,504 sequences were used, with 5% for validation and 2,784 for testing. The maximum interaction history length considered is 300 interactions, with a maximum token cost of 45,000.

Key Insights and Ideas

  • Unified Approach for Diverse Tasks: The paper demonstrates that a single multimodal model can effectively handle a range of educational tasks (knowledge recommendation, tracing, time cost prediction, user answer prediction), moving beyond task-specific solutions.
  • Compression is Key for Efficiency: Compressing student interaction histories is a highly effective strategy for reducing VRAM usage and computational costs, making LMMs more practical for educational applications.
  • Generative Framework for Educational Tasks: Reframing educational tasks into a generative format using LMMs allows for more flexible and nuanced responses, especially for tasks like user answer prediction.
  • Model Size Matters for Generative Tasks: While discriminative tasks (like knowledge tracing) are less sensitive to model size, generative tasks (like knowledge recommendation and user answer prediction) benefit significantly from larger language models.

Key Results

  • Efficiency Gains: UniEDU achieves approximately a 3x increase in computational efficiency compared to fully fine-tuned models, with significant reductions in VRAM usage for both training and inference. For example, UniEDU-5B uses 9.5GB for inference, compared to 31.5GB for Qwen2-VL-2B and 43.9GB for Qwen2-VL-7B.
  • Performance: UniEDU demonstrates strong performance across the four educational tasks. It outperforms task-specific baselines in Knowledge Recommendation, Time Cost Prediction, and User Answer Prediction, achieving gains of up to 30% and 20% respectively.
  • Knowledge Tracing: While competitive in knowledge tracing, UniEDU slightly lags behind specialized models like extraKT and reKT, which are optimized for discriminative tasks. However, UniEDU's strength lies in its generalization across different scenarios.
  • Compression Impact: Varying the number of compression tokens shows a slight degradation in performance for Knowledge Recommendation and Tracing, but User Answer Prediction benefits from additional context. Overall, the compression offers substantial efficiency gains with minimal performance loss.

Limitations and Future Work

  • Generalizability Beyond Mathematics: UniEDU's generalizability to other subjects and task types (e.g., essay grading) has not been explored.
  • Compression Trade-offs: While effective, the compression strategy might introduce minor performance drops in complex generative tasks.
  • Latent Learner Traits: The interaction-history-based profile modeling may overlook crucial latent traits like motivation or learning style, suggesting that incorporating richer signals could enhance personalization.
  • Future Work: The authors suggest exploring UniEDU's applicability to other subjects and task types, further optimizing the compression strategy, and incorporating richer learner signals for improved personalization.

Summary updated

BibTeX
@inproceedings{chu2025uniedu,
  author = {Zhendong Chu and Jian Xie and Shen Wang and Zichao Wang and Qingsong Wen},
  title = {{UniEDU: Toward Unified and Efficient Large Multimodal Models for Educational Tasks}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2025},
}
Figure