arXiv subjects

Ayan Majumdar

Publications and source records attributed to Ayan Majumdar.

7 recordsLinked to original sources

arXiv2026

Can Foundation Models Moderate Online Content? Evaluating Instruction- vs. Example-Driven Policy Operationalization

The growing complexity of content moderation policies presents a critical challenge for their consistent operationalization. While foundation models possess the basic capabilities needed to confront this challenge, whether they can reliably moderate online content remains an unanswered question. In this paper, we systematically compare two competing paradigms for Vision-Language Model (VLM) guidance: an instruction-driven approach where models reason from policy precepts, and an example-driven approach where they generalize from prior precedents. We ground this investigation in ModerationBench, a new benchmark of 4,000 manually annotated, in-the-wild posts from the Bluesky platform. Our experiments reveal that foundation models can substantially outperform Bluesky's deployed moderation system, nearly tripling its $F_1$ score (0.60 vs. 0.22) on Random Posts in the benchmark, with both instruction- and example-driven paradigms achieving comparable peak effectiveness. Our findings thus chart a path toward reliable and adaptable policy operationalization at scale.

cs.CL↗

arXiv2026

Characterizing Bluesky Content Moderation Service: From Automation of Service to Landscape of Harms

Empirical research on content moderation is fundamentally constrained by the opaque deployment of moderation systems on major social media platforms. To this end, the recent emergence of decentralized platforms with transparent, public moderation logs presents an unprecedented opportunity for independent audits. In this work, we leverage this architectural transparency to conduct the first large-scale audit of the default moderation system on Bluesky, the Bluesky Moderation Service (BMS). Analyzing its 10.6M moderation labels from 2025, we investigate three foundational aspects: (i) its mechanism (the degree of automation versus human oversight), (ii) its efficacy (accuracy in detecting harms), and (iii) its purpose (the landscape of harms it identifies). Our findings reveal a human-AI collaborative system where labels for sexual and graphic content are applied automatically in seconds, while nuanced and high stakes labels require more human oversight, taking hours or days. Through a manual annotation study, we find the BMS operates with high precision (0.837), but struggles with low recall (0.222), with our annotators identifying 4.5$\times$ more harmful content than the moderation system in a random sample. Finally, unsupervised clustering of the most frequently applied labeled posts uncovers detected harms ranging from hostility in discourse toward protected groups to the spread of sexually explicit and other graphic content. Our work offers a look into the operational realities of a deployed moderation system, providing a concrete data-driven foundation for designing more effective and transparent moderation systems.

cs.CY↗

arXiv2026

Evaluating LLMs for Demographic-Targeted Social Bias Detection: A Comprehensive Benchmark Study

Large-scale web-scraped text corpora used to train general-purpose AI models often contain harmful demographic-targeted social biases, creating a regulatory need for data auditing and developing scalable bias-detection methods. Although prior work has investigated biases in text datasets and related detection methods, these studies remain narrow in scope. They typically focus on a single content type (e.g., hate speech), cover limited demographic axes, overlook biases affecting multiple demographics simultaneously, and analyze limited techniques. Consequently, practitioners lack a holistic understanding of the strengths and limitations of recent large language models (LLMs) for automated bias detection. In this study, we conduct a comprehensive benchmark study on English texts to assess the ability of LLMs in detecting demographic-targeted social biases. To align with regulatory requirements, we frame bias detection as a multi-label task of detecting targeted identities using a demographic-focused taxonomy. We then systematically evaluate models across scales and techniques, including prompting, in-context learning, and fine-tuning. Using twelve datasets spanning diverse content types and demographics, our study demonstrates the promise of fine-tuned smaller models for scalable detection. However, our analyses also expose persistent gaps across demographic axes and multi-demographic targeted biases, underscoring the need for more effective and scalable detection frameworks.

cs.CL↗

arXiv2026

From Universal to Individualized Actionability: Revisiting Personalization in Algorithmic Recourse

Algorithmic recourse aims to provide actionable recommendations that enable individuals to change unfavorable model outcomes, and prior work has extensively studied properties such as efficiency, robustness, and fairness. However, the role of personalization in recourse remains largely implicit and underexplored. While existing approaches incorporate elements of personalization through user interactions, they typically lack an explicit definition of personalization and do not systematically analyze its downstream effects on other recourse desiderata. In this paper, we formalize personalization as individual actionability, characterized along two dimensions: hard constraints that specify which features are individually actionable, and soft, individualized constraints that capture preferences over action values and costs. We operationalize these dimensions within the causal algorithmic recourse framework, adopting a pre-hoc user-prompting approach in which individuals express preferences via rankings or scores prior to the generation of any recourse recommendation. Through extensive empirical evaluation, we investigate how personalization interacts with key recourse desiderata, including validity, cost, and plausibility. Our results highlight important trade-offs: individual actionability constraints, particularly hard ones, can substantially degrade the plausibility and validity of recourse recommendations across amortized and non-amortized approaches. Notably, we also find that incorporating individual actionability can reveal disparities in the cost and plausibility of recourse actions across socio-demographic groups. These findings underscore the need for principled definitions, careful operationalization, and rigorous evaluation of personalization in algorithmic recourse.

cs.LG↗

arXiv2025

A Causal Framework to Measure and Mitigate Non-binary Treatment Discrimination

Fairness studies of algorithmic decision-making systems often simplify complex decision processes, such as bail or loan approvals, into binary classification tasks. However, these approaches overlook that such decisions are not inherently binary (e.g., approve or not approve bail or loan); they also involve non-binary treatment decisions (e.g., bail conditions or loan terms) that can influence the downstream outcomes (e.g., loan repayment or reoffending). In this paper, we argue that non-binary treatment decisions are integral to the decision process and controlled by decision-makers and, therefore, should be central to fairness analyses in algorithmic decision-making. We propose a causal framework that extends fairness analyses and explicitly distinguishes between decision-subjects' covariates and the treatment decisions. This specification allows decision-makers to use our framework to (i) measure treatment disparity and its downstream effects in historical data and, using counterfactual reasoning, (ii) mitigate the impact of past unfair treatment decisions when automating decision-making. We use our framework to empirically analyze four widely used loan approval datasets to reveal potential disparity in non-binary treatment decisions and their discriminatory impact on outcomes, highlighting the need to incorporate treatment decisions in fairness assessments. Moreover, by intervening in treatment decisions, we show that our framework effectively mitigates treatment discrimination from historical data to ensure fair risk score estimation and (non-binary) decision-making processes that benefit all stakeholders.

cs.LG↗

arXiv2022

Do Invariances in Deep Neural Networks Align with Human Perception?

An evaluation criterion for safe and trustworthy deep learning is how well the invariances captured by representations of deep neural networks (DNNs) are shared with humans. We identify challenges in measuring these invariances. Prior works used gradient-based methods to generate identically represented inputs (IRIs), ie, inputs which have identical representations (on a given layer) of a neural network, and thus capture invariances of a given network. One necessary criterion for a network's invariances to align with human perception is for its IRIs look 'similar' to humans. Prior works, however, have mixed takeaways; some argue that later layers of DNNs do not learn human-like invariances (\cite{jenelle2019metamers}) yet others seem to indicate otherwise (\cite{mahendran2014understanding}). We argue that the loss function used to generate IRIs can heavily affect takeaways about invariances of the network and is the primary reason for these conflicting findings. We propose an adversarial regularizer on the IRI generation loss that finds IRIs that make any model appear to have very little shared invariance with humans. Based on this evidence, we argue that there is scope for improving models to have human-like invariances, and further, to have meaningful comparisons between models one should use IRIs generated using the regularizer-free loss. We then conduct an in-depth investigation of how different components (eg architectures, training losses, data augmentations) of the deep learning pipeline contribute to learning models that have good alignment with humans. We find that architectures with residual connections trained using a (self-supervised) contrastive loss with $\ell_p$ ball adversarial data augmentation tend to learn invariances that are most aligned with humans. Code: \url{github.com/nvedant07/Human-NN-Alignment}.

cs.CV↗

arXiv2022

Don't Throw it Away! The Utility of Unlabeled Data in Fair Decision Making

Decision making algorithms, in practice, are often trained on data that exhibits a variety of biases. Decision-makers often aim to take decisions based on some ground-truth target that is assumed or expected to be unbiased, i.e., equally distributed across socially salient groups. In many practical settings, the ground-truth cannot be directly observed, and instead, we have to rely on a biased proxy measure of the ground-truth, i.e., biased labels, in the data. In addition, data is often selectively labeled, i.e., even the biased labels are only observed for a small fraction of the data that received a positive decision. To overcome label and selection biases, recent work proposes to learn stochastic, exploring decision policies via i) online training of new policies at each time-step and ii) enforcing fairness as a constraint on performance. However, the existing approach uses only labeled data, disregarding a large amount of unlabeled data, and thereby suffers from high instability and variance in the learned decision policies at different times. In this paper, we propose a novel method based on a variational autoencoder for practical fair decision-making. Our method learns an unbiased data representation leveraging both labeled and unlabeled data and uses the representations to learn a policy in an online process. Using synthetic data, we empirically validate that our method converges to the optimal (fair) policy according to the ground-truth with low variance. In real-world experiments, we further show that our training approach not only offers a more stable learning process but also yields policies with higher fairness as well as utility than previous approaches.

stat.ML↗