ICLE++: Modeling Fine-Grained Traits for Holistic Essay Scoring

Introduces ICLE++, a new annotated corpus of persuasive student essays that addresses critical limitations of the dominant ASAP benchmark in Automated Essay Scoring research. Unlike ASAP — used by virtually all recent AES models but limited to U.S. grade 7–10 native-English essays — ICLE++ provides both holistic scores and fine-grained trait-specific annotations, enabling evaluation of cross-corpus generalization, multi-trait scoring, and cross-prompt scoring. The authors demonstrate that models trained on trait-specific annotations transfer better across corpora than those trained on holistic scores alone, making ICLE++ a foundational resource for the next generation of AES research.

Key Contributions

Implications for AI in Education

ICLE++ strengthens the Automated Essay Scoring research ecosystem by providing a second major annotated corpus that can test whether AES advances generalize beyond ASAP. This connects directly to Assessment Validity concerns: if models trained on one corpus fail on another, their scores are not valid measures of writing quality. The trait-specific annotations also enable more diagnostically useful AES — rather than a single score, systems can report strengths and weaknesses across specific writing dimensions, supporting Formative Assessment and connecting to psychometric frameworks like Item Response Theory.

The work complements Psyscore Essay Scoring ZPD Feedback's psychometrically-aware trait scoring and Choi Anchor Aes Prompting 2025's finding that prompting-based AES with anchor papers approaches human reliability. While Choi et al. show what prompting can achieve holistically, ICLE++ provides the annotated data needed to evaluate whether trait-level scoring generalizes — a question also explored by Self Referential L2 Writing LLM Assessment for L2 learners. The cross-corpus generalizability question likewise connects to AI Scoring Language Bias Physics's finding that AI scoring systematically underestimates linguistically diverse students, and to Aiawe Automated Writing Evaluation's demonstration that open-source models can achieve strong AES performance on specific datasets.

Connected Concepts

Connected Articles

Citation

Shengjie Li, Vincent Ng (2026). ICLE++: Modeling Fine-Grained Traits for Holistic Essay Scoring. arXiv:2607.27671. cs.CL.