Model fairness is becoming important in class-incremental learning for Trustworthy AI. While accuracy has been a central focus in class-incremental learning, fairness has been relatively understudied. However, naively using all the samples of the current task for training results in unfair catastrophic forgetting for certain sensitive groups including classes. We theoretically analyze that forgetting occurs if the average gradient vector of the current task data is in an "opposite direction" compared to the average gradient vector of a sensitive group, which means their inner products are negative. We then propose a fair class-incremental learning framework that adjusts the training weights of current task samples to change the direction of the average gradient vector and thus reduce the forgetting of underperforming groups and achieve fairness. For various group fairness measures, we formulate optimization problems to minimize the overall losses of sensitive groups while minimizing the disparities among them. We also show the problems can be solved with linear programming and propose an efficient Fairness-aware Sample Weighting (FSW) algorithm. Experiments show that FSW achieves better accuracy-fairness tradeoff results than state-of-the-art approaches on real datasets. The source code is released at https://github.com/jyparkkr/FSW.

In the context of continual learning, acquiring new knowledge while maintaining previous knowledge presents a significant challenge. Existing methods often use experience replay techniques that store a small portion of previous task data for training. In experience replay approaches, data augmentation has emerged as a promising strategy to further improve the model performance by mixing limited previous task data with sufficient current task data. However, we theoretically and empirically analyze that training with mixed samples from random sample pairs may harm the knowledge of previous tasks and cause greater catastrophic forgetting. We then propose GradMix, a robust data augmentation method specifically designed for mitigating catastrophic forgetting in class-incremental learning. GradMix performs gradient-based selective mixup using a class-based criterion that mixes only samples from helpful class pairs and not from detrimental class pairs for reducing catastrophic forgetting. Our experiments on various real datasets show that GradMix outperforms data augmentation baselines in accuracy by minimizing the forgetting of previous knowledge.

Post-hoc explanations are widely used to justify, contest, and review automated decisions in high-stakes domains such as lending, employment, and healthcare. Among these methods, SHAP is often treated as providing a reliable account of which features mattered for an individual prediction and is routinely used to support recourse, oversight, and accountability. In practice, however, SHAP explanations can differ substantially across repeated runs, even when the individual, prediction task, and trained model are held fixed. We conceptualize and name this phenomenon explanation multiplicity: the existence of multiple, internally valid but substantively different explanations for the same decision. Explanation multiplicity poses a normative challenge for responsible AI deployment, as it undermines expectations that explanations can reliably identify the reasons for an adverse outcome. We present a comprehensive methodology for characterizing explanation multiplicity in post-hoc feature attribution methods, disentangling sources arising from model training and selection versus stochasticity intrinsic to the explanation pipeline. Furthermore, whether explanation multiplicity is surfaced depends on how explanation consistency is measured. Commonly used magnitude-based metrics can suggest stability while masking substantial instability in the identity and ordering of top-ranked features. To contextualize observed instability, we derive and estimate randomized baseline values under plausible null models, providing a principled reference point for interpreting explanation disagreement. Across datasets, model classes, and confidence regimes, we find that explanation multiplicity is widespread and persists even under highly controlled conditions, including high-confidence predictions. Thus explanation practices must be evaluated using metrics and baselines aligned with their intended societal role.

As large language models (LLMs) are increasingly deployed in real-world applications, ensuring their fair responses across demographics has become crucial. Despite many efforts, an ongoing challenge is hidden bias: LLMs appear fair under standard evaluations, but can produce biased responses outside those evaluation settings. In this paper, we identify framing – differences in how semantically equivalent prompts are expressed (e.g., "A is better than B" vs. "B is worse than A") – as an underexplored contributor to this gap. We first introduce the concept of "framing disparity" to quantify the impact of framing on fairness evaluation. By augmenting fairness evaluation benchmarks with alternative framings, we find that (1) fairness scores vary significantly with framing and (2) existing debiasing methods improve overall (i.e., frame-averaged) fairness, but often fail to reduce framing-induced disparities. To address this, we propose a framing-aware debiasing method that encourages LLMs to be more consistent across framings. Experiments demonstrate that our approach reduces overall bias and improves robustness against framing disparities, enabling LLMs to produce fairer and more consistent responses.

Facts change over time, making it essential for Large Language Models (LLMs) to handle time-sensitive factual knowledge accurately and reliably. Although factual Time-Sensitive Question-Answering (TSQA) tasks have been widely developed, existing benchmarks often face manual bottlenecks that limit scalable and comprehensive TSQA evaluation. To address this issue, we propose TDBench, a new benchmark that systematically constructs TSQA pairs by harnessing temporal databases and database techniques, such as temporal functional dependencies, temporal SQL, and temporal joins. We also introduce a new evaluation metric called time accuracy, which assesses the validity of time references in model explanations alongside traditional answer accuracy for a more fine-grained TSQA evaluation. Extensive experiments on contemporary LLMs show how TDBench enables scalable and comprehensive TSQA evaluation while reducing the reliance on human labor, complementing current TSQA evaluation approaches that largely center on Wikipedia/Wikidata by enabling LLM evaluation on application-specific data.

Aligning large language models to human preferences is crucial for real-world deployment but frequently incurs an alignment tax, leading to the catastrophic forgetting of pre-trained general capabilities. While previous works primarily frame this problem as an optimization or architectural challenge, the inherent characteristics of preference data that drive this degradation remain largely underexplored. In this paper, we propose BALIGN, a balanced data selection strategy that explicitly mitigates catastrophic forgetting while optimizing alignment efficacy. Through theoretical and empirical analyses of the preference optimization gradient, we identify three key data-centric features that dictate parameter drift: the reference model's log-probability margin, the token length difference between chosen and rejected responses, and the TF-IDF similarity to general capability corpora. By aggregating these orthogonal features into a unified composite risk score, BALIGN systematically filters out high-risk preference samples that disrupt intrinsic model parameters or provide minimal alignment utility. Extensive experiments on standard human preference datasets demonstrate that BALIGN strongly preserves foundational capabilities without compromising alignment gains, consistently achieving the optimal Pareto frontier with minimal computational overhead.

Large Language Models (LLMs) offer a promising solution to complement traditional teaching and address global teacher shortages that affect hundreds of millions of children, but they fail to provide grade-appropriate responses for students at different educational levels. We introduce a framework for finetuning LLMs to generate age-appropriate educational content across six grade levels, from lower elementary to adult education. Our framework successfully adapts explanations to match students' comprehension capacities without sacrificing factual correctness. This approach integrates seven established readability metrics through a clustering method and builds a comprehensive dataset for grade-specific content generation. Evaluations across multiple datasets with 208 human participants demonstrate substantial improvements in grade-level alignment, achieving a 35.64 percentage point increase compared to prompt-based methods while maintaining response accuracy. AI-assisted learning tailored to different grade levels has the potential to advance educational engagement and equity.

Local feature-based explanations are a key component of the XAI toolkit. These explanations compute feature importance values relative to an ``interpretable'' feature representation. In tabular data, feature values themselves are often considered interpretable. This paper examines the impact of data engineering choices on local feature-based explanations. We demonstrate that simple, common data engineering techniques, such as representing age with a histogram or encoding race in a specific way, can manipulate feature importance as determined by popular methods like SHAP. Notably, the sensitivity of explanations to feature representation can be exploited by adversaries to obscure issues like discrimination. While the intuition behind these results is straightforward, their systematic exploration has been lacking. Previous work has focused on adversarial attacks on feature-based explainers by biasing data or manipulating models. To the best of our knowledge, this is the first study demonstrating that explainers can be misled by standard, seemingly innocuous data engineering techniques.

We study model confidence calibration in class-incremental learning, where models learn from sequential tasks with different class sets. While existing works primarily focus on accuracy, maintaining calibrated confidence has been largely overlooked. Unfortunately, most post-hoc calibration techniques are not designed to work with the limited memories of old-task data typical in class-incremental learning, as retaining a sufficient validation set would be impractical. Thus, we propose T-CIL, a novel temperature scaling approach for class-incremental learning without a validation set for old tasks, that leverages adversarially perturbed exemplars from memory. Directly using exemplars is inadequate for temperature optimization, since they are already used for training. The key idea of T-CIL is to perturb exemplars more strongly for old tasks than for the new task by adjusting the perturbation direction based on feature distance, with the single magnitude determined using the new-task validation set. This strategy makes the perturbation magnitude computed from the new task also applicable to old tasks, leveraging the tendency that the accuracy of old tasks is lower than that of the new task. We empirically show that T-CIL significantly outperforms various baselines in terms of calibration on real datasets and can be integrated with existing class-incremental learning techniques with minimal impact on accuracy.

Generative models must ensure both privacy and fairness for Trustworthy AI. While these goals have been pursued separately, recent studies propose to combine existing privacy and fairness techniques to achieve both goals. However, naively combining these techniques can be insufficient due to privacy-fairness conflicts, where a sample in a minority group may be represented in ways that support fairness, only to be suppressed for privacy. We demonstrate how these conflicts lead to adverse effects, such as privacy violations and unexpected fairness-utility tradeoffs. To mitigate these risks, we propose PFGuard, a generative framework with privacy and fairness safeguards, which simultaneously addresses privacy, fairness, and utility. By using an ensemble of multiple teacher models, PFGuard balances privacy-fairness conflicts between fair and private training stages and achieves high utility based on ensemble learning. Extensive experiments show that PFGuard successfully generates synthetic data on high-dimensional data while providing both DP guarantees and convergence in fair generative modeling.

Multimodal models often over-rely on dominant modalities, failing to achieve optimal performance. While prior work focuses on modifying training objectives or optimization procedures, data-centric solutions remain underexplored. We propose MIDAS, a novel data augmentation strategy that generates misaligned samples with semantically inconsistent cross-modal information, labeled using unimodal confidence scores to compel learning from contradictory signals. However, this confidence-based labeling can still favor the more confident modality. To address this within our misaligned samples, we introduce weak-modality weighting, which dynamically increases the loss weight of the least confident modality, thereby helping the model fully utilize weaker modality. Furthermore, when misaligned features exhibit greater similarity to the aligned features, these misaligned samples pose a greater challenge, thereby enabling the model to better distinguish between classes. To leverage this, we propose hard-sample weighting, which prioritizes such semantically ambiguous misaligned samples. Experiments on multiple multimodal classification benchmarks demonstrate that MIDAS significantly outperforms related baselines in addressing modality imbalance.

Fair visual recognition has become critical for preventing demographic disparity. A major cause of model unfairness is the imbalanced representation of different groups in training data. Recently, several works aim to alleviate this issue using generated data. However, these approaches often use generated data to obtain similar amounts of data across groups, which is not optimal for achieving high fairness due to different learning difficulties and generated data qualities across groups. To address this issue, we propose a novel adaptive sampling approach that leverages both real and generated data for fairness. We design a bilevel optimization that finds the optimal data sampling ratios among groups and between real and generated data while training a model. The ratios are dynamically adjusted considering both the model's accuracy as well as its fairness. To efficiently solve our non-convex bilevel optimization, we propose a simple approximation to the solution given by the implicit function theorem. Extensive experiments show that our framework achieves state-of-the-art fairness and accuracy on the CelebA and ImageNet People Subtree datasets. We also observe that our method adaptively relies less on the generated data when it has poor quality. Our work shows the importance of using generated data together with real data for improving model fairness.

Large language models (LLMs) have achieved unprecedented performances in various applications, yet evaluating them is still challenging. Existing benchmarks are either manually constructed or are automatic, but lack the ability to evaluate the thought process of LLMs with arbitrary complexity. We contend that utilizing existing relational databases based on the entity-relationship (ER) model is a promising approach for constructing benchmarks as they contain structured knowledge that can be used to question LLMs. Unlike knowledge graphs, which are also used to evaluate LLMs, relational databases have integrity constraints that can be used to better construct complex in-depth questions and verify answers: (1) functional dependencies can be used to pinpoint critical keywords that an LLM must know to properly answer a given question containing certain attribute values; and (2) foreign key constraints can be used to join relations and construct multi-hop questions, which can be arbitrarily long and used to debug intermediate answers. We thus propose ERBench, which uses these integrity constraints to convert any database into an LLM benchmark. ERBench supports continuous evaluation as databases change, multimodal questions, and various prompt engineering techniques. In our experiments, we construct LLM benchmarks using databases of multiple domains and make an extensive comparison of contemporary LLMs. We show how ERBench can properly evaluate any LLM by not only checking for answer correctness, but also effectively verifying the rationales by looking for the right keywords.

Continuous machine learning pipelines are common in industrial settings where models are periodically trained on data streams. Unfortunately, concept drifts may occur in data streams where the joint distribution of the data X and label y, P(X, y), changes over time and possibly degrade model accuracy. Existing concept drift adaptation approaches mostly focus on updating the model to the new data possibly using ensemble techniques of previous models and tend to discard the drifted historical data. However, we contend that explicitly utilizing the drifted data together leads to much better model accuracy and propose Quilt, a data-centric framework for identifying and selecting data segments that maximize model accuracy. To address the potential downside of efficiency, Quilt extends existing data subset selection techniques, which can be used to reduce the training data without compromising model accuracy. These techniques cannot be used as is because they only assume virtual drifts where the posterior probabilities P(y|X) are assumed not to change. In contrast, a key challenge in our setup is to also discard undesirable data segments with concept drifts. Quilt thus discards drifted data segments and selects data segment subsets holistically for accurate and efficient model training. The two operations use gradient-based scores, which have little computation overhead. In our experiments, we show that Quilt outperforms state-of-the-art drift adaptation and data selection baselines on synthetic and real datasets.

Large Language Models (LLMs) often struggle with requests related to information retrieval and data manipulation that frequently arise in real-world scenarios under multiple conditions. In this paper, we demonstrate that leveraging tabular structures in LLM interactions, is more effective than utilizing other structures for handling prevalent requests that operate over factual data. Through comprehensive evaluations across various scenarios and request types, we show that providing tabular structures yields a 40.29% average performance gain along with better robustness and token efficiency. Through attention-value analysis, we discover that tables help LLMs better locate relevant information, explaining these improvements. Beyond tables and text, we evaluate whether (1) blending structuredness within text, such as providing templates or fixing the order of attributes, and (2) other representative structures, such as knowledge graphs and JSON are helpful. We observe that utilizing tables offers the best balance between efficiency and effectiveness. The method remains robust to task complexity and adapts to unstructured sources through text-to-table conversion. Overall, we highlight the untapped potential of tabular representations for future LLM applications.

We study the problem of robust data augmentation for regression tasks in the presence of noisy data. Data augmentation is essential for generalizing deep learning models, but most of the techniques like the popular Mixup are primarily designed for classification tasks on image data. Recently, there are also Mixup techniques that are specialized to regression tasks like C-Mixup. In comparison to Mixup, which takes linear interpolations of pairs of samples, C-Mixup is more selective in which samples to mix based on their label distances for better regression performance. However, C-Mixup does not distinguish noisy versus clean samples, which can be problematic when mixing and lead to suboptimal model performance. At the same time, robust training has been heavily studied where the goal is to train accurate models against noisy data through multiple rounds of model training. We thus propose our data augmentation strategy RC-Mixup, which tightly integrates C-Mixup with multi-round robust training methods for a synergistic effect. In particular, C-Mixup improves robust training in identifying clean data, while robust training provides cleaner data to C-Mixup for it to perform better. A key advantage of RC-Mixup is that it is data-centric where the robust model training algorithm itself does not need to be modified, but can simply benefit from data mixing. We show in our experiments that RC-Mixup significantly outperforms C-Mixup and robust training baselines on noisy data benchmarks and can be integrated with various robust training methods.

Fine-tuning is becoming widely used for leveraging the power of pre-trained foundation models in new downstream tasks. While there are many successes of fine-tuning on various tasks, recent studies have observed challenges in the generalization of fine-tuned models to unseen distributions (i.e., out-of-distribution; OOD). To improve OOD generalization, some previous studies identify the limitations of fine-tuning data and regulate fine-tuning to preserve the general representation learned from pre-training data. However, potential limitations in the pre-training data and models are often ignored. In this paper, we contend that overly relying on the pre-trained representation may hinder fine-tuning from learning essential representations for downstream tasks and thus hurt its OOD generalization. It can be especially catastrophic when new tasks are from different (sub)domains compared to pre-training data. To address the issues in both pre-training and fine-tuning data, we propose a novel generalizable fine-tuning method LEVI (Layer-wiseEnsemble of differentVIews), where the pre-trained model is adaptively ensembled layer-wise with a small task-specific model, while preserving its efficiencies. By combining two complementing models, LEVI effectively suppresses problematic features in both the fine-tuning data and pre-trained model and preserves useful features for new tasks. Broad experiments with large language and vision models show that LEVI greatly improves fine-tuning generalization via emphasizing different views from fine-tuning data and pre-trained features.