PsyScore: A Psychometrically-Aware Framework for Trait-Adaptive Essay Scoring and ZPD-Scaffolded Feedback

Automated Essay Scoring (AES) systems often suffer from a disconnect between scoring and feedback. While neural models can predict scores, they typically act as black boxes with limited educational interpretability. Conversely, LLM-based feedback is often generic and fails to account for a student's specific proficiency level. This paper asks whether modeling student ability within a shared psychometric latent space can unify interpretable scoring with effective, scaffolded feedback.

The authors propose PsyScore, a framework that bridges diagnostic assessment and instructional support. The system consists of three primary modules:

1. Trait-Adaptive Neural IRT Scorer: This module embeds the Graded Partial Credit Model (GPCM) into a neural architecture. By mapping essay representations to a latent ability parameter (theta), the model provides interpretable scores grounded in psychometric theory.
2. ZPD-Scaffolded Feedback Generator: This module uses the diagnosed ability parameter to determine the appropriate pedagogical strategy—ranging from explicit correction for novices to Socratic questioning for advanced learners—within the student's Zone of Proximal Development (ZPD).
3. Multi-Perspective Evaluation: The framework evaluates feedback quality not just through lexical metrics, but through pairwise preference judgments by LLM judges and simulated student revision tasks.

Experiments on the ASAP++ dataset demonstrate that PsyScore achieves competitive, often superior, scoring performance compared to existing state-of-the-art models. Crucially, the integration of IRT parameters allows the system to generate feedback that is more pedagogically aligned with the learner's proficiency. The ablation studies confirm that the IRT component is essential for both the accuracy of the scoring and the diagnostic utility of the feedback.