Qianyi Fang
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Qianyi Fang has a total of 6 co-authors. Here are the top co-authors by number of shared publications:
Wenhe Liu
2 shared publications- Jie Si Yang
The University of Utah, Salt Lake City, UT, USA
1 shared publication - Minjae Rhee
University of Illinois Urbana–Champaign
1 shared publication - Wenhe Liu
School of Computer Science, Carnegie Mellon University, Pittsburgh, PA 15213, USA
1 shared publication Zixuan Guo
1 shared publication- Zixuan Guo
New York University
1 shared publication
Latest Publications
- Open Access
- 13 November 2026
- Published by Elsevier in SSRN Electronic Journal
Beyond Prediction Accuracy: Evaluating Fairness, Interpretability, and Educational Actionability in Student Early-warning Systems
Student early-warning systems (EWS) are usually chosen by predictive accuracy, yet the decisions they support depend on who is missed, who is flagged without need, whether explanations can be trusted, and whether warnings point to conditions that schools can change. We introduce FIA-Audit, a capacity-aware framework that jointly evaluates the Fairness, Interpretability, and educational Actionability of EWS. FIA-Audit organizes predictors into five domains, measures group-wise errors and calibration at a fixed intervention capacity, quantifies the stability and cross-model agreement of local explanations, and defines two actionability measures: the actionable attribution share (AAS +) and the lever-based recourse feasibility rate (RFR). Applied to a longitudinally linked HSLS:09 cohort of 14,563 U.S. ninth graders followed to 2016, the audit compares an indicator rule, logistic regression, random forest, and gradient boosting at two warning times and four feature regimes under school-grouped cross-validation. Machine-learning models raise the AUC from 0.737 to 0.826, but at a 20% capacity they flag 79.3% of eventual dropouts in the lowest socioeconomic quintile and 24.2% in the highest, while flagging 34.4% versus 2.2% of non-dropouts. Removing demographic attributes increases the over-flagging of Black students. A lever-centric model restricted to prior history and malleable indicators costs 0.009 AUC while raising AAS + from 0.496 to 0.633, recourse feasibility from 42.8% to 50.0%, and cross-model explanation agreement from 0.470 to 0.551. A crossfitted capacity-aware equal-recall allocation narrows the racial recall gap from 0.438 to 0.015 at a cost of 1.4 points of recall.
- 13 November 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessBeyond Prediction Accuracy: Evaluating Fairness, Interpretability, and Educational Actionability in Student Early-warning Systems
Student early-warning systems (EWS) are usually chosen by predictive accuracy, yet the decisions they support depend on who is missed, who is flagged without need, whether explanations can be trusted, and whether warnings point to conditions that schools can change. We introduce FIA-Audit, a capacity-aware framework that jointly evaluates the Fairness, Interpretability, and educational Actionability of EWS. FIA-Audit organizes predictors into five domains, measures group-wise errors and calibration at a fixed intervention capacity, quantifies the stability and cross-model agreement of local explanations, and defines two actionability measures: the actionable attribution share (AAS +) and the lever-based recourse feasibility rate (RFR). Applied to a longitudinally linked HSLS:09 cohort of 14,563 U.S. ninth graders followed to 2016, the audit compares an indicator rule, logistic regression, random forest, and gradient boosting at two warning times and four feature regimes under school-grouped cross-validation. Machine-learning models raise the AUC from 0.737 to 0.826, but at a 20% capacity they flag 79.3% of eventual dropouts in the lowest socioeconomic quintile and 24.2% in the highest, while flagging 34.4% versus 2.2% of non-dropouts. Removing demographic attributes increases the over-flagging of Black students. A lever-centric model restricted to prior history and malleable indicators costs 0.009 AUC while raising AAS + from 0.496 to 0.633, recourse feasibility from 42.8% to 50.0%, and cross-model explanation agreement from 0.470 to 0.551. A crossfitted capacity-aware equal-recall allocation narrows the racial recall gap from 0.438 to 0.015 at a cost of 1.4 points of recall.
- Open Access
- 13 November 2026
- Published by Elsevier in SSRN Electronic Journal
From AI Evaluation to Educational Practice: Developing an Educator-guided Framework for Fair and Responsible AI use in Schools
AI essay scorers and writing assistants are entering classrooms faster than schools can judge whether they are fit for a given use. Aggregate agreement statistics tell developers how close a model is to human raters on average; educators need to know for which students, on which assignments, and with how much teacher review a tool can be used responsibly. We present CLEAR, an educator-guided framework that turns educator-declared requirements-use case, error tolerance, priority student groups and review capacity-into auditable evidence and deployment decisions. CLEAR couples a layered audit of scores and diagnostic feedback with a peer-matched equity audit, group-conditional risk-controlled deferral (GC-RCD) that certifies with finitesample guarantees the error rate of released AI scores within every designated student group, teacher-seeded calibration for new prompts, and readiness cards stating the teacher review each use case requires. On 5,531 PERSUADE 2.0 essays with demographic data and expert discourse annotations, four scoring systems reach a quadratic weighted kappa of 0.76-0.83. The conventional subgroup check reports English language learners (ELLs) as overscored (SMD = +0.19), whereas comparison with peers who wrote the same prompt and received the same human score shows that they are underscored (-0.10). F1-tuned structural feedback tells ELLs who already wrote a counterclaim to add one 33 percentage points more often than matched peers; a within-group cap on unwarranted criticism shrinks the average largest-group gap from 19.6 to 7.7 points. At a 2% tolerance for non-adjacent errors, GC-RCD reviews 25.7% of essays and keeps every group within tolerance, while global thresholding leaves 3.6% of ELLs with unreviewed severe errors. Fifty teacher-scored essays remove 72% of the score bias on a new prompt.
- 13 November 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessFrom AI Evaluation to Educational Practice: Developing an Educator-guided Framework for Fair and Responsible AI use in Schools
AI essay scorers and writing assistants are entering classrooms faster than schools can judge whether they are fit for a given use. Aggregate agreement statistics tell developers how close a model is to human raters on average; educators need to know for which students, on which assignments, and with how much teacher review a tool can be used responsibly. We present CLEAR, an educator-guided framework that turns educator-declared requirements-use case, error tolerance, priority student groups and review capacity-into auditable evidence and deployment decisions. CLEAR couples a layered audit of scores and diagnostic feedback with a peer-matched equity audit, group-conditional risk-controlled deferral (GC-RCD) that certifies with finitesample guarantees the error rate of released AI scores within every designated student group, teacher-seeded calibration for new prompts, and readiness cards stating the teacher review each use case requires. On 5,531 PERSUADE 2.0 essays with demographic data and expert discourse annotations, four scoring systems reach a quadratic weighted kappa of 0.76-0.83. The conventional subgroup check reports English language learners (ELLs) as overscored (SMD = +0.19), whereas comparison with peers who wrote the same prompt and received the same human score shows that they are underscored (-0.10). F1-tuned structural feedback tells ELLs who already wrote a counterclaim to add one 33 percentage points more often than matched peers; a within-group cap on unwarranted criticism shrinks the average largest-group gap from 19.6 to 7.7 points. At a 2% tolerance for non-adjacent errors, GC-RCD reviews 25.7% of essays and keeps every group within tolerance, while global thresholding leaves 3.6% of ELLs with unreviewed severe errors. Fifty teacher-scored essays remove 72% of the score bias on a new prompt.
- Open Access
- 29 August 2026
- Published by Elsevier in SSRN Electronic Journal
Calibrating Trust in Generative AI Tutors: TRANSFER, a Control-Referenced Learning Analytics Method for Detecting Student Over-Reliance
Generative AI tutors can raise performance while students have access to assistance yet fail to improve, or even impair, performance once that assistance is removed. This creates a measurement problem for learning analytics: high assisted scores can indicate either effective scaffolding or dependency. We introduce TRANSFER, a controlreferenced sequential learning analytics method that detects this form of over-reliance without equating AI use frequency or assisted success with risk. The method first constructs a Control-Referenced Transfer Deficit (CRTD) by learning the independent-test score expected for a student under a no-AI reference condition and comparing that expectation with the student's realized unaided performance. A sequential risk model then predicts high CRTD from variables available at two deployment points: TRANSFER-Early, before the current AI-assisted practice episode using baseline and prior-session state, and TRANSFER-Post, after assisted practice but before the independent test. Finally, a budget-aware policy converts predicted risk into selective assignment of pedagogical guardrails. We evaluate the method on a verified 300-session, 92-student, 42-class reproducibility subset of the public randomized high-school mathematics dataset of Bastani et al. The evaluation uses student-grouped five-fold cross-validation, class-grouped robustness checks, strong static and sequential baselines, ablations, threshold and temporal sensitivity analyses, alternative reference models, grade-wise checks, and cluster bootstrap intervals. TRANSFER-Early achieves AUROC 0.676 and Brier 0.180; TRANSFER-Post achieves AUPRC 0.403 and ECE 0.083. At a 25% alert budget, Early warnings identify sessions with a 45.7% observed high-deficit rate, a 1.83× lift over the 25% base rate. Removing temporal state reduces AUPRC from 0.403 to 0.368 and worsens ECE from 0.083 to 0.110. In the same randomized dataset, the pedagogically guarded GPT Tutor arm shows a 0.031 lower mean CRTD than GPT Base in the local subset, while the published full-sample randomized trial establishes that guardrails largely eliminate the independent-learning penalty observed with unrestricted GPT assistance. TRANSFER therefore supplies an outcome-grounded detector and an operational bridge from risk estimation to targeted trust calibration.
- 29 August 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessCalibrating Trust in Generative AI Tutors: TRANSFER, a Control-Referenced Learning Analytics Method for Detecting Student Over-Reliance
Generative AI tutors can raise performance while students have access to assistance yet fail to improve, or even impair, performance once that assistance is removed. This creates a measurement problem for learning analytics: high assisted scores can indicate either effective scaffolding or dependency. We introduce TRANSFER, a controlreferenced sequential learning analytics method that detects this form of over-reliance without equating AI use frequency or assisted success with risk. The method first constructs a Control-Referenced Transfer Deficit (CRTD) by learning the independent-test score expected for a student under a no-AI reference condition and comparing that expectation with the student's realized unaided performance. A sequential risk model then predicts high CRTD from variables available at two deployment points: TRANSFER-Early, before the current AI-assisted practice episode using baseline and prior-session state, and TRANSFER-Post, after assisted practice but before the independent test. Finally, a budget-aware policy converts predicted risk into selective assignment of pedagogical guardrails. We evaluate the method on a verified 300-session, 92-student, 42-class reproducibility subset of the public randomized high-school mathematics dataset of Bastani et al. The evaluation uses student-grouped five-fold cross-validation, class-grouped robustness checks, strong static and sequential baselines, ablations, threshold and temporal sensitivity analyses, alternative reference models, grade-wise checks, and cluster bootstrap intervals. TRANSFER-Early achieves AUROC 0.676 and Brier 0.180; TRANSFER-Post achieves AUPRC 0.403 and ECE 0.083. At a 25% alert budget, Early warnings identify sessions with a 45.7% observed high-deficit rate, a 1.83× lift over the 25% base rate. Removing temporal state reduces AUPRC from 0.403 to 0.368 and worsens ECE from 0.083 to 0.110. In the same randomized dataset, the pedagogically guarded GPT Tutor arm shows a 0.031 lower mean CRTD than GPT Base in the local subset, while the published full-sample randomized trial establishes that guardrails largely eliminate the independent-learning penalty observed with unrestricted GPT assistance. TRANSFER therefore supplies an outcome-grounded detector and an operational bridge from risk estimation to targeted trust calibration.
- Open Access
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 138096-138116
Closed-Loop Audio Neuroregulation for ADHD: Graph-Based Attention Decoding and Rhythm Entrainment
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 138096-138116
Open AccessClosed-Loop Audio Neuroregulation for ADHD: Graph-Based Attention Decoding and Rhythm Entrainment
- Open Access
HARLA-ED: Resolving Information Asymmetry and Enhancing Algorithmic Symmetry in Intelligent Educational Assessment via Hybrid Reinforcement Learning
Conventional educational assessments enforce a rigid and symmetrical framework of identical question sequences upon a learner population inherently defined by asymmetry in cognitive capabilities and knowledge profiles. This mismatch results in inefficient measurement, where the uniform distribution of difficulty fails to mirror the heterogeneous nature of student learning. We address these topological and informational asymmetries through HARLA-ED, a hybrid framework combining deep knowledge modeling with intelligent question selection. The system integrates hierarchical cognitive graph networks to map the structural symmetries of concept dependencies while tracking evolving knowledge states across multiple time scales. By capturing both immediate working-memory constraints and long-term retention patterns, the model resolves the temporal asymmetry between learning and forgetting rates. A hierarchical reinforcement learning agent then orchestrates an assessment strategy through three decision levels: high-level planning determines diagnostic objectives, mid-level control sequences question types, and low-level actions select specific items. Crucially, the agent employs information-theoretic reward functions designed to restore distributional symmetry in assessment outcomes, ensuring demographic parity and minimizing algorithmic bias. Empirical results demonstrate a 47.5% average reduction in assessment duration compared to standard computer-adaptive tests while preserving measurement accuracy. The system successfully adapts to varying proficiency levels, effectively bridging the information asymmetry between the testing system and the learner’s true latent state.
Open AccessHARLA-ED: Resolving Information Asymmetry and Enhancing Algorithmic Symmetry in Intelligent Educational Assessment via Hybrid Reinforcement Learning
Conventional educational assessments enforce a rigid and symmetrical framework of identical question sequences upon a learner population inherently defined by asymmetry in cognitive capabilities and knowledge profiles. This mismatch results in inefficient measurement, where the uniform distribution of difficulty fails to mirror the heterogeneous nature of student learning. We address these topological and informational asymmetries through HARLA-ED, a hybrid framework combining deep knowledge modeling with intelligent question selection. The system integrates hierarchical cognitive graph networks to map the structural symmetries of concept dependencies while tracking evolving knowledge states across multiple time scales. By capturing both immediate working-memory constraints and long-term retention patterns, the model resolves the temporal asymmetry between learning and forgetting rates. A hierarchical reinforcement learning agent then orchestrates an assessment strategy through three decision levels: high-level planning determines diagnostic objectives, mid-level control sequences question types, and low-level actions select specific items. Crucially, the agent employs information-theoretic reward functions designed to restore distributional symmetry in assessment outcomes, ensuring demographic parity and minimizing algorithmic bias. Empirical results demonstrate a 47.5% average reduction in assessment duration compared to standard computer-adaptive tests while preserving measurement accuracy. The system successfully adapts to varying proficiency levels, effectively bridging the information asymmetry between the testing system and the learner’s true latent state.