Generative AI can produce programming feedback at scale, yet ensuring that the feedback aligns with instructors' pedagogical goals remains difficult. Instructors often cannot enumerate validation criteria in advance because such criteria emerge while inspecting concrete examples and correcting misaligned feedback.
ValiAssist introduces a validation‑by‑demonstration approach: the system watches instructors revise AI‑generated feedback, infers reusable validation logic, and applies it across other student submissions while allowing instructors to inspect and refine the logic on the fly. This iterative process makes clear which judgments generalize and which remain case‑specific.
In a comparative study with 16 instructors, ValiAssist enabled faster creation and refinement of validation criteria. Feedback aligned with pedagogical intent rose to 91% (vs. 57% without the tool), and the accuracy of correcting problematic criteria more than doubled from 32% to 77%.
These findings suggest that validation by demonstration helps turn individual case judgments into explicit, reusable logic, progressively clarifying the boundaries of their applicability.
Review