In this paper we present a systematic study of social bias in small- to mid-scale Large Language Models (LLMs), focusing on gender, religion, and race. Using our SALT (Social Appropriateness in LLM Text) dataset, we explore two bias categories—Theoretical and Practical. Theoretical bias covers General Debate and Positioned Debate while practical bias includes Career Advice, Personal Advice, and Resume Generation. We quantify bias using win-rate gaps in general debate, and negative-role assignments in positioned debate. For Practical bias, we anonymize model outputs to remove explicit demographic cues and use DeepSeek-R1 as an automated evaluator, measuring outcome disparities across groups. We also examine systemic issues in LLM-based evaluation including evaluation bias, positional bias, and length bias and validate our findings through human annotation. Our results show consistent disadvantages for White, Christian, and male-associated outputs across multiple tasks. Larger models often amplify these disparities, highlighting that scale does not guarantee fairness.

This paper presents a comparative analysis of Large Language Models (LLMs) and traditional Optical Character Recognition (OCR) systems on Urdu newspapers, addressing challenges posed by complex multi-column layouts, low-resolution scans, and the stylistic variability of the Nastaliq script. To handle these challenges, we fine-tune YOLOv11x models for article- and column-level text block extraction and train a SwinIR-based super-resolution module that enhances image quality for downstream text recognition, improving accuracy by an average of 50%. We further introduce the Urdu Newspaper Benchmark (UNB), a manually annotated dataset for Urdu OCR comprising 829 paragraph images with a total of 9,982 sentences. Using UNB and the OpenITI corpus, we conduct a systematic comparison between traditional CNN+RNN-based OCR systems and modern LLMs, presenting detailed insertion, deletion, and substitution error analyses alongside character-level confusion patterns. We find that Gemini-2.5-Pro achieves the best performance on UNB (WER 0.133), while fine-tuning GPT-4o on just 500 in-domain samples yields a 6.13% absolute WER improvement, demonstrating the adaptability of LLMs to low-resource, morphologically complex scripts like Urdu. The UNB dataset and fine-tuned models are publicly available at https://github.com/paper-seven/UrduOCR.

Children increasingly have access to Large Language Models (LLMs), which may expose them to responses that are developmentally inappropriate or require age-sensitive safety, guidance, and boundaries. Existing LLM safety evaluations largely focus on general harmful-content avoidance and do not explicitly target child-facing safety. We introduce KIDBench, a benchmark for evaluating child-facing LLM safety for ages 7-11 using a LLM-as-a-Judge rubric grounded in developmental-psychology. KIDBench contains realistic child queries across ten categories, with single-turn prompts and multi-turn child-actor simulations. We compare no-cues prompts with no child context, implicit-cues prompts that suggest a child speaker, and explicit age instructions. Implicit-cues improve scores by 8.6-46.8% over no-cue, while explicit age provides an additional 9.9-30.4% improvement over implicit-cues. Cross-lingual and cultural evaluations show uneven safety behavior across languages and country contexts. Multi-turn simulations show peak quality drops of up to 0.959 points on the 1-5 scale. We also introduce KIDGuardLlama, a child-safety evaluator, and KIDLlama, a child-safe response model. Code, data, and evaluation resources are available at https://github.com/MichiganNLP/kidbench.

Fairness benchmarks such as BBQ have become the de facto standard for fairness evaluation across major model families. We argue that these benchmarks are too easy to saturate to support that role: training Qwen 2.5 7B Base with Group Relative Policy Optimization (GRPO) on a single BBQ example lifts mean BBQ accuracy from 79.9 to 92.0 across all 11 social-bias categories, closing of the gap to the large-scale RLHF counterpart, and the effect generalizes across model families. A cross-conditioning analysis shows the improvement is carried by the reasoning traces generated by the model, not by the output head; one example suffices to elicit a category-agnostic “missing evidence” reasoning pattern. Strikingly, placing that single example in context as a one-shot demonstration—with no weight update—likewise reaches comparable BBQ accuracy and produces reasoning a fairness reward model rates as equally fair, confirming that the benchmark is saturable without gradient-based alignment. We argue that BBQ-style multiple-choice abstention benchmarks measure a single structural cue, and a model that solves them does not thereby become fair. We call for evaluation suites that cover a broader spectrum of fairness alignment.

This paper introduces Kahaani, a multimodal, co-creative storytelling system that leverages Generative Artificial Intelligence, designed for children to address the challenge of sustaining engagement to foster educational narrative experiences. Here we define co-creative as a collaborative creative process in which both the child and Kahaani contribute to the generation of the story. The system combines Large Language Model (LLM), Text-to-Speech (TTS), Text-to-Music (TTM), and Text-to-Video (TTV) generation to produce a rich, immersive, and accessible storytelling experience. The system grounds the co-creation process in two classical storytelling framework, Freytag’s Pyramid and Propp’s Narrative Functions. The main goals of Kahaani are: (1) to help children improve their English skills, (2) to teach important life lessons through story morals, and (3) to help them understand how stories are structured, all in a fun and engaging way. We present evaluations for each AI component used, along with a user study involving three parent–child pairs to assess the overall experience and educational value of the system.

This paper presents a novel methodology for generating synthetic Preference Optimization (PO) datasets using multi-model workflows. We evaluate the effectiveness and potential of these workflows in automating and enhancing the dataset generation process. PO dataset generation requires two modules: (1) response evaluation, and (2) response generation. In the response evaluation module, the responses from Large Language Models (LLMs) are evaluated and ranked - a task typically carried out by human annotators that we automate using LLMs. We assess the response evaluation module in a 2 step process. In step 1, we assess LLMs as evaluators using three distinct prompting strategies. In step 2, we apply the winning prompting strategy to compare the performance of LLM-as-a-Judge, LLMs-as-a-Jury, and LLM Debate. Our evaluation shows that GPT-4o-as-a-Judge is more consistent across all datasets. For the response generation module, we use the identified LLM evaluator configuration and compare different configurations of the LLM Feedback Loop. We use the win rate to determine the best multi-model configuration for generation. Experimenting with various configurations, we find that the LLM Feedback Loop, with Llama as the generator and Gemma as the reviewer, achieves a notable 71.8% and 73.8% win rate over single-model Llama and Gemma, respectively. After identifying the best configurations for both modules, we generate our PO datasets using the above pipeline.

This paper presents a comprehensive evaluation of Urdu Automatic Speech Recognition (ASR) models. We analyze the performance of three ASR model families: Whisper, MMS, and Seamless-M4T using Word Error Rate (WER), along with a detailed examination of the most frequent wrong words and error types including insertions, deletions, and substitutions. Our analysis is conducted using two types of datasets, read speech and conversational speech. Notably, we present the first conversational speech dataset designed for benchmarking Urdu ASR models. We find that seamless-large outperforms other ASR models on the read speech dataset, while whisper-large performs best on the conversational speech dataset. Furthermore, this evaluation highlights the complexities of assessing ASR models for low-resource languages like Urdu using quantitative metrics alone and emphasizes the need for a robust Urdu text normalization system. Our findings contribute valuable insights for developing robust ASR systems for low-resource languages like Urdu.

This paper introduces UQA, a novel dataset for question answering and text comprehension in Urdu, a low-resource language with over 70 million native speakers. UQA is generated by translating the Stanford Question Answering Dataset (SQuAD2.0), a large-scale English QA dataset, using a technique called EATS (Enclose to Anchor, Translate, Seek), which preserves the answer spans in the translated context paragraphs. The paper describes the process of selecting and evaluating the best translation model among two candidates: Google Translator and Seamless M4T. The paper also benchmarks several state-of-the-art multilingual QA models on UQA, including mBERT, XLM-RoBERTa, and mT5, and reports promising results. For XLM-RoBERTa-XL, we have an F1 score of 85.99 and 74.56 EM. UQA is a valuable resource for developing and testing multilingual NLP systems for Urdu and for enhancing the cross-lingual transferability of existing models. Further, the paper demonstrates the effectiveness of EATS for creating high-quality datasets for other languages and domains. The UQA dataset and the code are publicly available at www.github.com/sameearif/UQA