A LLM-Powered Automatic Grading Framework with Human-Level Guidelines Optimization

Open-ended short-answer questions (SAGs) have been widely recognized as a powerful tool for providing deeper insights into learners' responses in the context of learning analytics (LA). However, SAGs often present challenges in practice due to the high grading workload and concerns about inconsistent assessments. With recent advancements in natural language processing (NLP), automatic short-answer grading (ASAG) offers a promising solution to these challenges. Despite this, current ASAG algorithms are often limited in generalizability and tend to be tailored to specific questions. In this paper, we propose a unified multi-agent ASAG framework, GradeOpt, which leverages large language models (LLMs) as graders for SAGs. More importantly, GradeOpt incorporates two additional LLM-based agents - the reflector and the refiner - into the multi-agent system. This enables GradeOpt to automatically optimize the original grading guidelines by performing self-reflection on its errors. Through experiments on a challenging ASAG task, namely the grading of pedagogical content knowledge (PCK) and content knowledge (CK) questions, GradeOpt demonstrates superior performance in grading accuracy and behavior alignment with human graders compared to representative baselines. Finally, comprehensive ablation studies confirm the effectiveness of the individual components designed in GradeOpt.

Augment with Care: Enhancing Graph Contrastive Learning with Selective Spectrum Perturbation

In recent years, Graph Contrastive Learning (GCL) has shown remarkable effectiveness in learning representations on graphs. As a component of GCL, good augmentation views are supposed to be invariant to the important information while discarding the unimportant part. Existing augmentation views with perturbed graph structures are usually based on random topology corruption in the spatial domain; however, from perspectives of the spectral domain, this approach may be ineffective as it fails to pose tailored impacts on the information of different frequencies, thus weakening the agreement between the augmentation views. By a preliminary experiment, we show that the impacts caused by spatial random perturbation are approximately evenly distributed among frequency bands, which may harm the invariance of augmentations required by contrastive learning frameworks. To address this issue, we argue that the perturbation should be selectively posed on the information concerning different frequencies. In this paper, we propose GASSER which poses tailored perturbation on the specific frequencies of graph structures in spectral domain, and the edge perturbation is selectively guided by the spectral hints. As shown by extensive experiments and theoretical analysis, the augmentation views are adaptive and controllable, as well as heuristically fitting the homophily ratios and spectrum of graph structures.