From Knowledge Concepts to Process-Based Skills: A Feedback-Driven Ability Induction Framework for Student Modeling

ACL ARR 2026 January Submission6097 Authors

Keywords: student modeling, cognitive diagnosis, knowledge tracing, educational data mining

Abstract:

Ability representations are central to student modeling because they connect item requirements to students' observed responses.
Most existing approaches use expert-defined knowledge concepts as ability units, which describe what content an item involves but often miss how it is solved when solving requires multi-step procedures.
We study a process-based alternative by inducing ability units from solution procedures.
Specifically, we treat large language model (LLM) reasoning traces as noisy process observations, and propose a closed-loop framework that extracts operation-level signals from structured LLM reasoning and consolidates them into stable, reusable process-based abilities.
To reduce noise and keep induced abilities aligned with the target items, we combine semantic consolidation with a global feasibility signal from downstream student modeling, and use a TextGrad-style discrete controller to refine generation constraints and consolidation hyperparameters.
Experiments on two benchmark datasets show that replacing expert concepts with induced process-based abilities improves student modeling performance and yields coherent ability structures with informative granularity and coverage.

Paper Type: Long

Research Area: NLP Applications

Research Area Keywords: educational applications, cognitive modeling, inductive reasoning

Contribution Types: NLP engineering experiment, Data analysis

Languages Studied: Chinese, English

Submission Number: 6097