Large Language Models (LLMs) have shown impressive capabilities across many language understanding tasks; however, their ability to perform numerical reasoning remains fragile, particularly outside English. This limitation is especially concerning for Hindi, a language with hundreds of millions of speakers but relatively little focused evaluation in numerically grounded reasoning. In this work, we present HI-REASO, a benchmark designed to systematically evaluate numerical reasoning in Hindi. The benchmark comprises six complementary task types and 3k instances drawn from educational materials, real-world problem formulations, and commonsense reasoning scenarios. We assess a range of state-of-the-art LLMs using Hindi-native prompting as well as cross-lingual prompting and cross-lingual chain-of-thought reasoning. Across all settings, models consistently underperform compared to human baselines, with the largest performance degradation observed when reasoning is conducted directly in Hindi. While cross-lingual strategies offer partial improvements, they do not close the gap. Our findings reveal fundamental shortcomings in current LLMs’ numerical reasoning abilities under Hindi-native conditions and motivate the development of models and training paradigms that better support linguistically diverse numerical reasoning. We publicly release the code and data at https://github.com/kraritt/hi-reaso.

The widespread use of the internet has made sharing information on social media more convenient. At the same time, it provides a platform for individuals with malicious intent to easily spread hateful content. Since many users prefer to communicate in their native language, detecting hate speech in Bengali poses a significant challenge. This study aims to identify Bengali hate speech on social media platforms. A shared task on Bengali hate speech detection was organized by the Second Bangla Language Processing Workshop (BLP). To tackle this task, we implemented five traditional machine learning models (LR, SVM, RF, NB, XGB), three deep learning models (CNN, BiLSTM, CNN+BiLSTM), and three transformer-based models (Bangla-BERT, m-BERT, XLM-R). Among all models, a weighted ensemble of transformer models achieved the best performance.Our approach ranked 3rd in Subtask 1A with a micro-F1 score of 0.734, 6th in Subtask 1B with 0.7315, and, after post-competition experiments, 4th in Subtask 1C with 0.735.