Research Papers

Backchannels are subtle listener responses, such as nods, smiles, or short verbal cues like "yes" or "uh-huh," which convey understanding and agreement in conversations. These signals provide feedback to speakers, improve the smoothness of interaction, and play a crucial role in developing human-like, responsive AI systems. However, the expression of backchannel behaviors is often significantly influenced by individual differences, operating across multiple scales: from instant dynamics such as response intensity (frame-level) to temporal patterns such as frequency and rhythm preferences (sequence-level). This presents a complex pattern recognition problem that contemporary emotion recognition methods have yet to fully address. Particularly, existing individualized methods in emotion recognition often operate at a single scale, overlooking the complementary nature of multi-scale behavioral cues. To address these challenges, we propose a novel Cascaded Multi-Scale Individual Standardization Network (CMIS-Net) that extracts individual-normalized backchannel features by removing person-specific neutral baselines from observed expressions. Operating at both frame and sequence levels, this normalization allows model to focus on relative changes from each person's baseline rather than absolute expression values. Furthermore, we introduce an implicit data augmentation module to address the observed training data distributional bias, improving model generalization. Comprehensive experiments and visualizations demonstrate that CMIS-Net effectively handles individual differences and data imbalance, achieving state-of-the-art performance in backchannel agreement detection.

Attribution Graphs and Causal Probing for Mechanistic Discovery and Bias Repair in Multimodal Generative Learning

Noor Islam S. Mohammad, Uluğ Bayazıt

We treat the internals of generative models as mechanistic objects rather than black boxes. We introduce \textbf{Attribution Graphs} (AGs), which extend GradCAM++ to circuit-level representations, and \textbf{Causal Probing}, a do-calculus intervention method for identifying causal latent structures, enabling detection and correction of spurious correlations, demographic biases, and misaligned decision circuits during training. We further propose the \textbf{Cognitive Alignment Score (CAS)}, quantifying agreement between model-internal representations and human concepts, a \textbf{saliency-first privacy mechanism} sharing only thresholded attribution nodes, a bias-aware regularizer aligning subgroup statistics, and a Reveal-to-Revise loop integrating attribution signals into parameter updates without separate fine-tuning. Evaluated on CelebA, FairFace, Jigsaw, and HateXplain, our method achieves \textbf{94.1\%} accuracy, \textbf{92.3\%} macro F1, \textbf{79.4\%} IoU-XAI, and \textbf{12.7} FID at 72--76\% adversarial robustness, while reducing subgroup disparity $Δ_{\mathrm{bias}}$ by \textbf{41\%}, demonstrating that mechanistic interpretability, fairness, and generative performance can be jointly optimized.

This paper presents a framework for provably fair AI that overcomes the limits of current bias mitigation methods by systematically removing all sensitive information and its proxies. Using ontology engineering in OWL 2 QL, it formally defines sensitive attributes and infers their proxies through logical reasoning, constructing a sigma algebra G that captures the full structure of biased patterns. Fair representations are then obtained via Delbaen Majumdar optimal transport, which generates variables independent of G while minimizing L2 distance to preserve accuracy. This guarantees true independence rather than mere decorrelation. By modeling bias as dependence between sigma algebras, compiling ontological knowledge into measurable structures, and using optimal transport as the unique fair transformation, the approach ensures complete fairness in tasks like loan approval, where proxies such as ZIP code reveal race. The result is a certifiable and mathematically grounded method for trustworthy AI.

humancompatible.detect: a Python Toolkit for Detecting Bias in AI Models

German M. Matilla, Jiri Nemecek, Illia Kryvoviaz +1

There is a strong recent emphasis on trustworthy AI. In particular, international regulations, such as the AI Act, demand that AI practitioners measure data quality on the input and estimate bias on the output of high-risk AI systems. However, there are many challenges involved, including scalability (MMD) and computability (Wasserstein-1) issues of traditional methods for estimating distances on measure spaces. Here, we present humancompatible.detect, a toolkit for bias detection that addresses these challenges. It incorporates two newly developed methods to detect and evaluate bias: maximum subgroup discrepancy (MSD) and subsampled $\ell_\infty$ distances. It has an easy-to-use API documented with multiple examples. humancompatible.detect is licensed under the Apache License, Version 2.0.

Accurate, fair, and explainable allocation of candidates to roles represents a fundamental challenge across multiple domains including corporate hiring, academic admissions, fellowship awards, and volunteer placement systems. Current state-of-the-art approaches suffer from semantic inflexibility, persistent demographic bias, opacity in decision-making processes, and poor scalability under dynamic policy constraints. We present GESA (Graph-Enhanced Semantic Allocation), a comprehensive framework that addresses these limitations through the integration of domain-adaptive transformer embeddings, heterogeneous self-supervised graph neural networks, adversarial debiasing mechanisms, multi-objective genetic optimization, and explainable AI components. Our experimental evaluation on large-scale international benchmarks comprising 20,000 candidate profiles and 3,000 role specifications demonstrates superior performance with 94.5% top-3 allocation accuracy, 37% improvement in diversity representation, 0.98 fairness score across demographic categories, and sub-second end-to-end latency. Additionally, GESA incorporates hybrid recommendation capabilities and glass-box explainability, making it suitable for deployment across diverse international contexts in industry, academia, and non-profit sectors.

Large language models (LLMs) are increasingly central to many applications, raising concerns about bias, fairness, and regulatory compliance. This paper reviews risks of biased outputs and their societal impact, focusing on frameworks like the EU's AI Act and the Digital Services Act. We argue that beyond constant regulation, stronger attention to competition and design governance is needed to ensure fair, trustworthy AI. This is a preprint of the Communications of the ACM article of the same title.

This paper offers a domain-mediated comparative review of 251 studies on public attitudes toward AI, published between 2011 and 2025. Drawing on a systematic literature review, we analyse how different factors including perceived benefits and concerns (or risks) shape public acceptance of - or resistance to - artificial intelligence across domains and use-cases, including healthcare, education, security, public administration, generative AI, and autonomous vehicles. The analysis highlights recurring patterns in individual, contextual, and technical factors influencing perception, while also tracing variations in institutional trust, perceived fairness, and ethical concerns. We show that the public perception in AI is shaped not only by technical design or performance but also by sector-specific considerations as well as imaginaries, cultural narratives, and historical legacies. This comparative approach offers a foundation for developing more tailored and context-sensitive strategies for responsible AI governance.

Existing deep learning models for chest radiology often neglect patient metadata, limiting diagnostic accuracy and fairness. To bridge this gap, we introduce MetaCheX, a novel multimodal framework that integrates chest X-ray images with structured patient metadata to replicate clinical decision-making. Our approach combines a convolutional neural network (CNN) backbone with metadata processed by a multilayer perceptron through a shared classifier. Evaluated on the CheXpert Plus dataset, MetaCheX consistently outperformed radiograph-only baseline models across multiple CNN architectures. By integrating metadata, the overall diagnostic accuracy was significantly improved, measured by an increase in AUROC. The results of this study demonstrate that metadata reduces algorithmic bias and enhances model generalizability across diverse patient populations. MetaCheX advances clinical artificial intelligence toward robust, context-aware radiographic disease detection.

EPT Benchmark: Evaluation of Persian Trustworthiness in Large Language Models

Mohammad Reza Mirbagheri, Mohammad Mahdi Mirkamali, Zahra Motoshaker Arani +3

Large Language Models (LLMs), trained on extensive datasets using advanced deep learning architectures, have demonstrated remarkable performance across a wide range of language tasks, becoming a cornerstone of modern AI technologies. However, ensuring their trustworthiness remains a critical challenge, as reliability is essential not only for accurate performance but also for upholding ethical, cultural, and social values. Careful alignment of training data and culturally grounded evaluation criteria are vital for developing responsible AI systems. In this study, we introduce the EPT (Evaluation of Persian Trustworthiness) metric, a culturally informed benchmark specifically designed to assess the trustworthiness of LLMs across six key aspects: truthfulness, safety, fairness, robustness, privacy, and ethical alignment. We curated a labeled dataset and evaluated the performance of several leading models - including ChatGPT, Claude, DeepSeek, Gemini, Grok, LLaMA, Mistral, and Qwen - using both automated LLM-based and human assessments. Our results reveal significant deficiencies in the safety dimension, underscoring the urgent need for focused attention on this critical aspect of model behavior. Furthermore, our findings offer valuable insights into the alignment of these models with Persian ethical-cultural values and highlight critical gaps and opportunities for advancing trustworthy and culturally responsible AI. The dataset is publicly available at: https://github.com/Rezamirbagheri110/EPT-Benchmark.

Predictive models are typically trained on historical data to predict future outcomes. While it is commonly assumed that training on more historical data would improve model performance and robustness, data distribution shifts over time may undermine these benefits. This study examines how expanding historical data training windows under covariate shifts (changes in feature distributions) and concept shifts (changes in feature-outcome relationships) affects the performance and algorithmic fairness of predictive models. First, we perform a simulation study to explore scenarios with varying degrees of covariate and concept shifts in training data. Absent distribution shifts, we observe performance gains from longer training windows though they reach a plateau quickly; in the presence of concept shift, performance may actually decline. Covariate shifts alone do not significantly affect model performance, but may complicate the impact of concept shifts. In terms of fairness, models produce more biased predictions when the magnitude of concept shifts differs across sociodemographic groups; for intersectional groups, these effects are more complex and not simply additive. Second, we conduct an empirical case study of student retention prediction, a common machine learning application in education, using 12 years of student records from 23 minority-serving community colleges in the United States. We find concept shifts to be a key contributor to performance degradation when expanding the training window. Moreover, model fairness is compromised when marginalized populations have distinct data distribution shift patterns from their peers. Overall, our findings caution against conventional wisdom that "more data is better" and underscore the importance of using historical data judiciously, especially when it may be subject to data distribution shifts, to improve model performance and fairness.

The use of inappropriate language -- such as outdated, exclusionary, or non-patient-centered terms -- medical instructional materials can significantly influence clinical training, patient interactions, and health outcomes. Despite their reputability, many materials developed over past decades contain examples now considered inappropriate by current medical standards. Given the volume of curricular content, manually identifying instances of inappropriate use of language (IUL) and its subcategories for systematic review is prohibitively costly and impractical. To address this challenge, we conduct a first-in-class evaluation of small language models (SLMs) fine-tuned on labeled data and pre-trained LLMs with in-context learning on a dataset containing approximately 500 documents and over 12,000 pages. For SLMs, we consider: (1) a general IUL classifier, (2) subcategory-specific binary classifiers, (3) a multilabel classifier, and (4) a two-stage hierarchical pipeline for general IUL detection followed by multilabel classification. For LLMs, we consider variations of prompts that include subcategory definitions and/or shots. We found that both LLama-3 8B and 70B, even with carefully curated shots, are largely outperformed by SLMs. While the multilabel classifier performs best on annotated data, supplementing training with unflagged excerpts as negative examples boosts the specific classifiers' AUC by up to 25%, making them most effective models for mitigating harmful language in medical curricula.

Large Language Models (LLMs) are widely used for downstream tasks such as tabular classification, where ensuring fairness in their outputs is critical for inclusivity, equal representation, and responsible AI deployment. This study introduces a novel approach to enhancing LLM performance and fairness through the concept of a dynamic validation set, which evolves alongside the test set, replacing the traditional static validation approach. We also propose an iterative algorithm, SMITE, to select optimal in-context examples, with each example set validated against its corresponding dynamic validation set. The in-context set with the lowest total error is used as the final demonstration set. Our experiments across four different LLMs show that our proposed techniques significantly improve both predictive accuracy and fairness compared to baseline methods. To our knowledge, this is the first study to apply dynamic validation in the context of in-context learning for LLMs.