Recognizing disinformation is a challenging task for humans and AI systems. News can be false, misleading, or harmful, and its interpretation often depends on the cultural context of the audience. However, existing datasets rarely account for these contextual and cultural differences, as they are typically not designed from the perspective of news consumers. To address this gap, in this paper, we present the Information Disorder (InDor) corpus, a multilingual dataset of news articles in English, Farsi, Italian, and Russian, annotated for information disorder detection and explanation. The corpus was developed through a participatory process involving contributors from diverse cultural and professional backgrounds, who engaged in data collection, annotation, and evaluation of Large Language Model (LLM) performance on the task. Our findings highlight that false and manipulated news manifest differently across cultural settings, and that current LLMs fail to adequately capture this complexity. This underscores the need for culturally aware computational approaches in the study of information disorder.

Bangla (Bengali), spoken by over 300 million people worldwide, presents unique tokenization challenges due to its rich morphological structure. Despite its global importance, Bangla remains underrepresented in mainstream NLP, particularly in low-resource tokenization research. Existing tokenizers such as Byte Pair Encoding (BPE), Unigram, WordPiece, and character-level were originally developed for high-resource languages. These tokenizers struggle with morphologically rich languages and show two main limitations: (i) unnecessary splitting of morphemes and (ii) fragmentation of compound words. Based on these limitations, we hypothesize that morphology-aware tokenization is necessary for handling Bangla effectively. To address this, we first introduce a benchmark that systematically evaluates how widely used tokenizers perform on morphology-related tasks. Motivated by these findings, we then propose BANGLATOKENIZER, a morphology-aware tokenizer that alleviates these limitations using a novel bond-strength framework inspired by molecular structures. Empirically, BANGLATOKENIZER yields consistent downstream improvements, achieving absolute accuracy gains of 2.8%-3.4%, macro-F1 gains of 2.5%-3.6%, and perplexity reductions of up to 20% across diverse language models and tasks. These improvements are accompanied by 5%-10% higher inference throughput, showing that morphology-aware tokenization improves both effectiveness and efficiency

Large Language Models (LLMs) are increasingly deployed in education, powering tutoring systems, automated feedback, and curriculum support. Most alignment techniques optimize toward a single aggregated notion of human preference, a mismatch for education, where legitimate disagreement is the norm. Students, parents, teachers, and policymakers often hold deeply conflicting views on assessment, discipline, curriculum, and institutional authority, reflecting genuine value pluralism rather than noise. Pluralistic alignment, which seeks to represent and navigate this diversity, offers a principled path forward. Despite growing interest, no benchmark evaluates how well LLMs handle value-laden disagreement in educational contexts. To address this, we introduce EQUAL (Eliciting QUality sAmples of pLuralistic alignment in education), comprising 14,270 samples—5,824 value-laden scenarios and 8,446 multiple-choice questions, curated from six sources, including over two decades of Phi Delta Kappa polling data, via a lexical filtering pipeline and human validation. EQUAL supports evaluation across Overton, Steerable, and Distributional paradigms. We benchmark nine LLMs under four approaches and additionally evaluate three frontier reasoning-capable models. No single method consistently satisfies pluralistic objectives. These findings provide a foundation for future research on pluralistic alignment in educational AI.