Literary Chinese reading comprehension is a challenging task for large language models (LLMs), requiring understanding of implicit meanings, contextual reasoning, and discourse-level integration in a low-resource setting. However, existing evaluations rely primarily on aggregate accuracy, which provides limited insight into model capabilities, error sources, and reliability under uncertainty. We propose DELit, a diagnostic evaluation framework that moves beyond aggregate metrics by introducing a hierarchical capability–error taxonomy, a Deep-MIRT-based ability estimation model, and a capability-decoupled residual analysis for behavioral diagnosis. Together, these components enable fine-grained interpretation of model performance at both capability and error-type levels. Experiments on mainstream LLMs reveal that (i) discourse-level understanding remains a consistent bottleneck despite strong performance on local semantic tasks; (ii) similar overall accuracy can correspond to substantially different capability profiles; and (iii) overconfident errors on difficult items concentrate on logically constrained reasoning, highlighting reliability risks that are not captured by standard evaluation metrics.

Multi-trait automated essay scoring extends holistic scoring by evaluating essays across multiple writing dimensions. In this framework, inter-trait consistency signals provide an effective way to improve prediction across traits. However, learning shared scoring signals may obscure trait-specific features, while enforcing consistency across traits can introduce negative transfer, hindering the model’s ability to distinguish individual traits. To solve these issues, we propose DPRES, a dual-perspective decoupled representation learning framework that disentangles consistency and specificity at both the feature and optimization levels. Specifically, we first propose a residual subspace decomposition strategy that enforces orthogonality between shared scoring features and trait-specific features, thereby disentangling shared scoring signals from trait-specific scoring signals in the latent representation space and enhancing the model’s ability to capture trait-specific scoring features. Based on the subspace decomposition strategy, we further design a consistency–specificity decoupled learning strategy to simultaneously preserve inter-trait correlations and enhance trait-level discrimination. Experiments on public datasets demonstrate that the proposed method consistently outperforms strong baselines.

Educational knowledge graph (EKG) is a critical component of intelligent tutoring systems that is structured around cognitive principles and provides support for interactive teaching. Most existing EKGs usually rely on simplistic relations, bind with single subjects, and lack integration with explicit learning objectives. In this paper, we introduce CogNet-KG, a novel and cognitively-structured large-scale knowledge graph for STEM learning. CogNet-KG models nearly 500 core concepts across five subjects with various cognitively-grounded relations corresponding to specific learning objectives, thereby encoding a rich cognitive schema for guiding more effective teaching. Based on this structure, we then construct a high-quality tutoring dialogue dataset CogDialogue-QA by leveraging adaptive instructional strategies. Additionally, we train CogTutor-LM, a specialized tutorial LLM that internalizes this structured pedagogical reasoning. Overall evaluation demonstrates that CogTutor-LM generates responses with significantly greater instructional coherence and more appropriate pedagogical guidance compared to baselines, validating the effectiveness of our graph-driven approach to fostering knowledge integration and stimulating students’ thinking. The datasets are publicly available at https://github.com/KCAIED/CogNet-KG.

Cross-prompt trait scoring task aims to learn generalizable scoring capabilities from source- prompt data, enabling automatic scoring across multiple dimensions on unseen essays. Existing research on cross-prompt trait essay scoring primarily focuses on improving model generalization by obtaining prompt-invariant representations. In this paper, we approach the research problem from a different perspective on invariance learning and propose a scoring-invariant learning objective. This objective encourages the model to focus on intrinsic information within the essay that reflects its quality during training, thereby learning generic scoring features. To further enhance the model’s ability to score across multiple dimensions, we introduce a trait feature extraction network based on routing gates into the scoring architecture and propose a trait consistency scoring objective to encourage the model to balance the diversity of trait-specific features with scoring consistency across traits when learning trait-specific essay features. Extensive experiments demonstrate the effectiveness of our approach, showing advantages in multi-trait scoring performance and achieving significant improvements with low-resource prompts.