LLM-Supported Hybrid Rubric Assessment in Multi-Stage Algebraic Block Coding Tasks: Design and Initial Evaluation

Assessment of multi-stage mathematics tasks requires explicit links between student work and evaluation criteria. This study presents a large language model (LLM)-supported hybrid rubric framework combining rule-based correctness checks, categorical evaluation of task-segment evidence, and separate qualitative feedback generation. Its task-specific profile distinguishes conceptual understanding, symbolic procedures, block-coding procedures, self-reported persistence, and a written explanation of tool usefulness. An initial evaluation involved 42 middle-school students undertaking a quadratic-equation task. Two experts independently scored student work without access to system scores; averaged ratings for 40 students yielded 200 paired rubric scores. System and expert scores showed moderate association (Pearson correlation = 0.632; Spearman rank correlation = 0.676), with a mean absolute error of 17.3 points and mean bias of +5.1 points. Learner-cluster bootstrap intervals accounted for scores nested within students. Excluding self-reported persistence yielded 160 pairs, a Pearson correlation of 0.669, and a mean absolute error of 16.8. Block-coding scores showed the strongest association with expert scores, while symbolic-procedure scores had the smallest absolute errors. Individual score differences were substantial. The study provides a task-specific design and an initial empirical reference for developing teacher-supervised hybrid assessment.