Aug 2026· Scientific Reports· Vol 16· 0 citations· 48 references
TL;DR
This work develops a mathematically robust framework of deep optimization with multiple objectives to ensure equity is maintained and carried forward across the tutoring lifecycle and achieves up to 67% lower learning gain disparity and 56% greater stability of dispersion in learning-gap distributions.
Abstract
The advancements of modern artificial intelligence in education (AIEd) systems have greatly improved prediction accuracy and personal pacing. Traditional intelligent tutoring systems have the highest achievable aggregate predictive accuracy, a value that often is rooted in historical biases, mis-represents engagement signals from advantaged learner groups, and leaves vulnerable learner groups out of the picture. To address these challenges, we develop a mathematically robust framework of deep optimization with multiple objectives to ensure equity is maintained and carried forward across the tutoring lifecycle. We propose 5 main components: (i) demographic sensitivity gradient encoding (DSGE) for measures and limits direct demographic influence by computing gradient-level sensitivities of the learning loss with respect to latent demographic embeddings; (ii) counterfactual equity replay networks (CERN) which guided by the DSGE signal, CERN models the learning process through an explicit structural causal model and uses offline counterfactual simulation to quantify fairness-sensitive trajectory differences under stated identification assumptions. (iii) The engage-weighted fairness attention fusion dynamically balances student persistence and fairness risk, in order to not let high-level engagements mask concerns for fairness. (iv) Pareto-adaptive equity-accuracy-engagement optimizer adapts objective weights on the simplex space through a meta-gradient optimization which promotes stable convergence across the competing objectives under the adopted training procedure. In (v) equity-preserving policy distillation and validation, the high-capacity multi-objective model is compressed to a light-weight student model while ensuring that the student model preserves the equity of the multi-objective model when deployed. Our framework is validated using three real-world, publicly available educational datasets viz., EdNet, ASSISTments, and open university learning analytics dataset. Empirical results indicate that our framework achieves up to 67% lower learning gain disparity and 56% greater stability of dispersion in learning-gap distributions, while maintaining learning accuracy within 1.2% of the unconstrained, accuracy-only baselines.
The surge in the use of artificial intelligence in hiring has created serious issues around algorithmic fairness and absence of bias in candidate assessment, as conventional AI-based interview systems are based on rigid rubrics that tend to reinforce demographic gaps. In order to overcome these issues, we introduce Fair- MARLTrans, a fairness-enforced multimodal adaptive reinforcement learning system for fair candidate assessment through virtual interviews. The model combines visual attributes (face landmarks, gaze, body posture), acoustic attributes (MFCC, prosody, spectral features), and textual attributes (BERT embeddings, sentiment) through a transformer-based attention mechanism to generate a single multimodal representation for end-to-end candidate profiling. Evaluation is framed as a Constrained Markov Decision Process (CMDP) with fairness-sensitive policy gradients imposing demographic parity, equalized opportunity, and individual fairness between protected groups[1]. We propose the Composite Interview Quality Index (CIQI) to simultaneously measure candidate competence and fairness compliance, with theoretical validity guarantees that provide convergence to fair policies. Experiment results on the CMU-MOSI dataset show that Fair-MARLTrans achieves better performance than static rubric baselines (15.14 ± 0.37 mean reward vs. 12.50 ± 0.85 for visual-only), standard reinforcement learning, and fairness-regularized models, with a fairness score of 0.97 and zero constraint violations. Ablation experiments show that multimodal fusion enhances performance by 18.5%, whereas attention mechanisms add 12.3% gains. Demographic sensitivity analysis for ten demographic subgroups holds firmly with fairness compliance (averaged score 0.73, p < 0.01)[2], confirmed by t-tests, ANOVA, and multiple comparison adjustment. This paper provides a theoretical basis for fairness-constrained adaptive assessment and a practical approach to transparent, fair AI-powered recruitment tools, with improved state-of-the-art in algorithmic recruitment.
D. M. J. Basha, D. A. Kumar· Journal of Intelligent Decis...· 0 citations
In order to tackle both of the described issues, the following research introduces a unique multi-stage solution, which includes a synthetic data generator based on a Conditional Tabular Generative Adversarial Network, used to balance the data, and a Transformer-based network for predicting the desired outcome. A CTGAN will synthesize a sufficient amount of high-fidelity examples for a minor class to ensure proper learning process while a modified transformer will recognize relevant relations between various professional features through its multi-head attention mechanism. For “Human-in-the-Loop” decision making, a SHAP layer is included. The Explainable Artificial Intelligence (XAI) module breaks down the intricate output generated by the Transformer model into detailed contribution scores, providing not only local reasoning for each evaluation but also global understanding of organizational productivity drivers. It has been empirically established that the GAN-Transformer-SHAP approach outperforms all baseline ensembles (random forest, XGBoost) and Bayesian techniques [10, 14] in terms of F1-score and AUC-ROC. At the end of the day, this study offers a reliable, interpretable, and unbiased path to adopting advanced AI into talent management and performance appraisal processes.
R. Wagh, R. Kharat· Engineering & Technology· 0 citations
The rapid integration of Artificial Intelligence (AI) and deep learning in higher education has enabled automated student performance prediction systems; however, these systems frequently inherit and amplify biases present in historical educational data, leading to inequitable outcomes for students across gender, socioeconomic, geographic, and linguistic dimensions. Existing machine learning approaches for academic performance prediction predominantly optimize for accuracy, largely disregarding fairness constraints, which risks systematically disadvantaging already marginalized learner groups. This paper proposes FairEduNet, a fairness-aware deep learning framework that integrates adversarial debiasing, fairness-regularized training, and SHAP-based explainability to simultaneously detect and mitigate bias in student performance prediction. The proposed architecture employs a multi-task neural network with a dedicated adversarial classifier that penalizes predictions correlated with protected attributes, supported by demographic parity and equalized odds constraints. Experimental evaluation on two publicly available educational datasets—the Open University Learning Analytics Dataset (OULAD) and the UCI Student Performance Dataset—demonstrates that FairEduNet achieves competitive predictive accuracy (F1: 0.87) while reducing demographic parity difference by 34% and equalized odds gap by 28% compared to state-of-the-art baselines. The framework further provides interpretable, per-group feature attribution via SHAP, enabling educators and policymakers to audit AI systems for fairness. The results establish that fairness and accuracy are not inherently at odds in educational prediction contexts, and that responsible AI design can produce both equitable and performant academic decision-support systems.
Training embodied foundation models typically requires massive-scale datasets and extensive computational resources, yet often suffers from three critical limitations: (1) inefficient sample utilization due to low-informative samples; (2) imbalanced gradient contributions across heterogeneous tasks; and (3) severe credit assignment problem in long-horizon planning, where trajectory-level rewards indiscriminately penalize all tokens. To address these issues, we propose an efficient training paradigm that achieves state-of-the-art average performance through strategic data selection and hierarchical policy optimization. Our approach consists of three synergistic stages. First, Rejection Sampling-based Fine-Tuning (RSFT) filters out low-informative samples to establish robust behavioral priors while preventing distributional collapse. Second, Iterative Rejection GRPO (IR-GRPO) employs task-specific queues stratified by difficulty to keep datasets balanced across reinforcement learning iterations, coupled with a hybrid reward mechanism for precise cross-task feedback. Third, to enhance long-horizon task planning, we introduce Trie-GRPO, a novel reinforcement learning algorithm based on action prefix trees, which enables step-level advantage estimation. This resolves the credit assignment problem by isolating intermediate correct decisions from downstream errors, while effectively balancing exploration efficiency and depth compared to conventional search trees. As a result, EmbodiedMind achieves a state-of-the-art average performance of 70.02% across 18 benchmarks, and significantly outperforms other embodied foundation models in long-horizon task planning accuracy. Our project will be released for reproducibility.
Fei-Fan Wang, Zong-Bing Zhang, Yu Zhang et al.· 0 citations
This study presents a context-aware Machine Learning (ML) framework for predicting student
performance and dropout in online learning environments affected by insurgency-related
disruptions. Unlike conventional models that assume stable learning conditions, the proposed
framework explicitly integrates environmental instability into its predictive architecture. The
system combines assessment records, institutional data, Learning Management System (LMS)
logs, and survey-derived contextual indicators within a unified analytical pipeline. Main variables
include forum participation, login frequency, time-on-task, submission regularity, and disruption
sensitive behavioral patterns linked to insecurity and displacement. To improve reliability under
unstable conditions, the framework incorporates normalization, missing-value imputation,
temporal alignment, and imbalance handling during preprocessing. Prediction is performed using
Decision Tree, Logistic Regression, Random Forest, and Support Vector Machine models to
enable robust comparative learning. A multi-objective optimization strategy balances predictive
accuracy, fairness, robustness, and interpretability, while SHAP-based explainable AI enhances
transparency in decision-making. The framework advances learning analytics through disruption
aware predictive intelligence for early identification of at-risk learners and supports resilient,
adaptive, and equitable online education in crisis-affected environments.
Abdulqadir Ahmad Bazway· International Journal of Com...· 0 citations
Graphical User Interface (GUI) agents powered by Multimodal Large Language Models (MLLMs) have shown strong potential for automating tasks across diverse digital environments, where reinforcement learning (RL) has become a dominant training paradigm. However, widely used methods such as Group Relative Policy Optimization (GRPO) suffer from reward-gradient misalignment, leading to inefficient and unstable optimization. Recent work addresses this issue by reformulating RL with verifiable rewards (RLVR) as contrastive or classification-based objectives, which improve stability by eliminating problematic gradient behaviors. Despite this progress, existing contrastive RLVR methods rely primarily on outcome-level supervision and fail to capture fine-grained differences in trajectory quality within the same outcome category. In this paper, we propose Length-Aware Contrastive Learning for GUI Agents (LACL-GUI), a contrastive RLVR framework that incorporates trajectory-level quality signals into policy optimization. LACL-GUI introduces structured preferences within both successful and failed trajectories, encouraging concise successful executions and differentiating failure quality based on divergence from successful trajectories, while preserving optimization stability. Experiments on GUI agent benchmarks show that LACL-GUI provides more effective learning signals and consistently improves agent performance over prior methods, highlighting the value of trajectory-level supervision in contrastive RLVR.
Chengyang Gu, Le Zhang, Jingbo Zhou et al.· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
