EUVIMEDEUVIMED entdecken
Open Access · CC BY-NC-ND 4.0

Process-aware programming assessment through temporal knowledge graph reasoning and LLM verification

Shuzhan Wu, Wenjun Zhou, Qiaofeng Zhang, Wentao Xiong

Discover Artificial Intelligence · 04.06.2026 · 6 (1)

DOI: 10.1007/s44163-026-01423-9 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Discover Artificial Intelligence, online erschienen am 04.06.2026. Verfasst von Shuzhan Wu und 3 weiteren Autor:innen. Die Arbeit nennt 56 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC-ND 4.0). Gefördert von: Project of Science and Technology Research of Hubei Provincial Department of Education, The Xiaogan City Natural Science Program Project, Research Project of Hubei Engineering University.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Automated programming assessment systems predominantly evaluate final code submissions while ignoring the problem-solving processes that reveal genuine learning, and large language model (LLM)-based feedback systems suffer from hallucinations due to the absence of structured domain knowledge grounding. This paper proposes a Process-AwareEvaluation Framework that is based on a combination of temporal knowledge graph reasoning with KG-augmented LLM verification, which can be used for the overall assessment of programming assignments. The proposed framework includes four modules that are tightly coupled and connected: Process-Aware Temporal Knowledge Graphs(PA-TKG), which models the evolution of code, concept coverage, and problem-solving activities as a timestamped heterogeneous graph structure; Composition-based Graph Convolutional Networks (CompGCN), which learn embeddings of relation-aware representations across different edge types; Contrastive Process Pattern Learning(CPPL)with a dual-encoder architecture that compares static code structure to dynamic development patterns to detect anomalous behavior; and KG-Guided Chain-of-Thought Verification (KG-CoTV), which constrains LLM reasoning against valid learning paths derived from a knowledge graph. The experiments conducted using two publicly available datasets demonstrated that the proposed framework achieved a process evaluation score of 0.912, indicating a 26.1 percent increase compared to the best performing baseline, as well as 62.8 percent higher anomaly detection accuracy and an 83.3 percent decrease in feedback hallucination rates. Ablation studies were also performed to verify the contributions of each module toward the overall framework performance, where it was found that the Process-Aware Temporal Knowledge Graphs (PA-TKG)provided the most significant representational basis. The core non-LLM pipeline (graph construction, embedding, anomaly detection, and constraint checking)takes less than 500 ms to process submissions; when combined with LLM inference for feedback generation, the total end-to-end latency averages 3.26 s on an NVIDIA A100 GPU, making the proposed framework suitable for near-real-time deployment in educational environments.

Förderung

  • Project of Science and Technology Research of Hubei Provincial Department of Education
  • The Xiaogan City Natural Science Program Project
  • Research Project of Hubei Engineering University

Zitieren

APA

Wu, S., Zhou, W., Zhang, Q., & Xiong, W. (2026). Process-aware programming assessment through temporal knowledge graph reasoning and LLM verification. Discover Artificial Intelligence, 6(1). https://doi.org/10.1007/s44163-026-01423-9

Vancouver

Wu S, Zhou W, Zhang Q, Xiong W. Process-aware programming assessment through temporal knowledge graph reasoning and LLM verification. Discover Artificial Intelligence. 2026;6(1). doi:10.1007/s44163-026-01423-9

BibTeX
@article{Wu2026,
  author = {Wu, Shuzhan and Zhou, Wenjun and Zhang, Qiaofeng and Xiong, Wentao},
  title = {Process-aware programming assessment through temporal knowledge graph reasoning and LLM verification},
  journal = {Discover Artificial Intelligence},
  year = {2026},
  volume = {6},
  number = {1},
  doi = {10.1007/s44163-026-01423-9}
}
Kostenlos starten →