Gagan Bhatia

Maria Kunilovskaya, Gagan Bhatia, Lisa Sophie Albertelli et al.

Human annotation is the empirical foundation of much NLP research, from dataset construction to model evaluation, but papers often leave unclear who produced the annotations and how the annotation process was controlled. We provide the first large-scale, task-level audit of human annotation reporting across major NLP venues, asking which annotation details are documented, which are missing, and how reporting varies across time, topic, venue, and intended use of human judgment. We introduce a unified taxonomy of annotation-reporting practices and validate an LLM-assisted extraction pipeline against Annotated-gold, a human-adjudicated gold standard of 41 papers and 72 annotation tasks, where the best model reaches human-comparable agreement with adjudicated labels, with Krippendorff's alpha of 0.606 versus 0.585 for human-human agreement. Using this pipeline, we construct Annotated-llm, a dataset covering ACL-venue papers from 2018-2025, with 2,667 extracted annotation tasks from 1,603 papers, and find that papers frequently report operational details such as recruitment strategies, annotator expertise, and annotation volume, but often omit details needed to assess annotation validity, including training, language proficiency, compensation, socio-demographics, adjudication, and agreement values, especially in model-evaluation studies. Our results show that annotation reporting in NLP has improved over time but remains uneven, and they establish a scalable framework and bare-minimum reporting recommendations for making human annotation more reliable, reproducible, and interpretable.

Firoj Alam, Gagan Bhatia, Sahinur Rahman Laskar et al.

While Large Language Models (LLMs) are increasingly adopted as automated judges for evaluating generated text, their outputs are often costly, and highly sensitive to prompt design, language, and aggregation strategies, severely, which limits reproducibility. To address these challenges, we propose \textbf{\textit{OmniScore}}, a family of complementary, deterministic learned metrics developed using small size ($<$1B) parameter models. OmniScore approximates LLM-judge behavior while preserving the low latency and consistency of traditional model-based scoring. We trained the models large-scale synthetic supervision ($\sim$564k instances, in \textbf{107 languages}) and evaluated using 8,617 manually annotated instances. The OmniScore family supports reliable, multi-dimensional scores across a variety of settings, including reference-based, source-grounded, and hybrid evaluations. We evaluate these models across question answering (QA), translation, and summarization in \textbf{6 languages}. Our results demonstrate that lightweight, deterministic learned metrics provide a highly practical and scalable alternative to frontier LLMs. Our models and datasets can be found at https://huggingface.co/collections/QCRI/omniscore

Fakhraddin Alwajih, Gagan Bhatia, Muhammad Abdul-Mageed

Recent advancements have significantly enhanced the capabilities of Multimodal Large Language Models (MLLMs) in generating and understanding image-to-text content. Despite these successes, progress is predominantly limited to English due to the scarcity of high quality multimodal resources in other languages. This limitation impedes the development of competitive models in languages such as Arabic. To alleviate this situation, we introduce an efficient Arabic multimodal assistant, dubbed Dallah, that utilizes an advanced language model based on LLaMA-2 to facilitate multimodal interactions. Dallah demonstrates state-of-the-art performance in Arabic MLLMs. Through fine-tuning six Arabic dialects, Dallah showcases its capability to handle complex dialectal interactions incorporating both textual and visual elements. The model excels in two benchmark tests: one evaluating its performance on Modern Standard Arabic (MSA) and another specifically designed to assess dialectal responses. Beyond its robust performance in multimodal interaction tasks, Dallah has the potential to pave the way for further development of dialect-aware Arabic MLLMs.

Gagan Bhatia, El Moatez Billah Nagoudi, Fakhraddin Alwajih et al.

Arabic Optical Character Recognition (OCR) and Handwriting Recognition (HWR) pose unique challenges due to the cursive and context-sensitive nature of the Arabic script. This study introduces Qalam, a novel foundation model designed for Arabic OCR and HWR, built on a SwinV2 encoder and RoBERTa decoder architecture. Our model significantly outperforms existing methods, achieving a Word Error Rate (WER) of just 0.80% in HWR tasks and 1.18% in OCR tasks. We train Qalam on a diverse dataset, including over 4.5 million images from Arabic manuscripts and a synthetic dataset comprising 60k image-text pairs. Notably, Qalam demonstrates exceptional handling of Arabic diacritics, a critical feature in Arabic scripts. Furthermore, it shows a remarkable ability to process high-resolution inputs, addressing a common limitation in current OCR systems. These advancements underscore Qalam's potential as a leading solution for Arabic script recognition, offering a significant leap in accuracy and efficiency.

Gagan Bhatia, Ahmad Muhammad Isa, Maxime Peyrard et al.

We present MultiTempBench, a multilingual temporal reasoning benchmark spanning three tasks, date arithmetic, time zone conversion, and temporal relation extraction across five languages (English, German, Chinese, Arabic, and Hausa) and multiple calendar conventions (Gregorian, Hijri, and Chinese Lunar). MultiTempBench contains $15,000$ examples built by translating $750$ curated English questions and expanding each into controlled date-format variants. We evaluate 20 LLMs and introduce the multilingual Date Fragmentation Ratio (mDFR), calibrated with human severity ratings, together with geometric-probing analyses of internal temporal representations. We find tokenisation quality of temporal artefacts is a resource-dependent bottleneck: in low-resource languages and rarer calendar formats, fragmentation disrupts Year/Month/Day separation and accuracy collapses, while high-resource settings are often robust to digit-level splitting. Beyond tokenisation, crossed mixed-effects regression shows that temporal linearity is the strongest predictor of temporal reasoning in high-resource languages, whereas fragmentation is the stronger predictor in low-resource languages. Code is available at: https://github.com/gagan3012/mtb

Sang Yun Kwon, Gagan Bhatia, El Moatez Billah Nagoud et al.

Recently, large language models (LLMs) fine-tuned to follow human instruction have exhibited significant capabilities in various English NLP tasks. However, their performance in grammatical error correction (GEC) tasks, particularly in non-English languages, remains significantly unexplored. In this paper, we delve into abilities of instruction fine-tuned LLMs in Arabic GEC, a task made complex due to Arabic's rich morphology. Our findings suggest that various prompting methods, coupled with (in-context) few-shot learning, demonstrate considerable effectiveness, with GPT-4 achieving up to $65.49$ F\textsubscript{1} score under expert prompting (approximately $5$ points higher than our established baseline). This highlights the potential of LLMs in low-resource settings, offering a viable approach for generating useful synthetic data for model training. Despite these positive results, we find that instruction fine-tuned models, regardless of their size, significantly underperform compared to fully fine-tuned models of significantly smaller sizes. This disparity highlights a substantial room for improvements for LLMs. Inspired by methods from low-resource machine translation, we also develop a method exploiting synthetic data that significantly outperforms previous models on two standard Arabic benchmarks. Our work sets new SoTA for Arabic GEC, with $72.19\%$ and $73.26$ F$_{1}$ on the 2014 and 2015 QALB datasets, respectively.

Gagan Bhatia, Hamdy Mubarak, Mustafa Jarrar et al.

LLMs are increasingly used for Islamic question answering, where ungrounded responses may carry serious religious consequences. Yet standard MCQ/MRC-style evaluations do not capture key real-world failure modes, notably free-form hallucinations and whether models appropriately abstain when evidence is lacking. To shed a light on this aspect we introduce ISLAMICFAITHQA, a 3,810-item bilingual (Arabic/English) generative benchmark with atomic single-gold answers, which enables direct measurement of hallucination and abstention. We additionally developed an end-to-end grounded Islamic modelling suite consisting of (i) 25K Arabic text-grounded SFT reasoning pairs, (ii) 5K bilingual preference samples for reward-guided alignment, and (iii) a verse-level Qur'an retrieval corpus of $\sim$6k atomic verses (ayat). Building on these resources, we develop an agentic Quran-grounding framework (agentic RAG) that uses structured tool calls for iterative evidence seeking and answer revision. Experiments across Arabic-centric and multilingual LLMs show that retrieval improves correctness and that agentic RAG yields the largest gains beyond standard RAG, achieving state-of-the-art performance and stronger Arabic-English robustness even with a small model (i.e., Qwen3 4B). We will make the experimental resources and datasets publicly available for the community.

Sang Yun Kwon, Gagan Bhatia, El Moatez Billah Nagoudi et al.

Intent detection and slot filling are critical tasks in spoken and natural language understanding for task-oriented dialog systems. In this work we describe our participation in the slot and intent detection for low-resource language varieties (SID4LR; Aepli et al. (2023)). We investigate the slot and intent detection (SID) tasks using a wide range of models and settings. Given the recent success of multitask-prompted finetuning of large language models, we also test the generalization capability of the recent encoder-decoder model mT0 (Muennighoff et al., 2022) on new tasks (i.e., SID) in languages they have never intentionally seen. We show that our best model outperforms the baseline by a large margin (up to +30 F1 points) in both SID tasks

Gagan Bhatia, Ife Adebara, AbdelRahim Elmadany et al.

We describe our contribution to the SemEVAl 2023 AfriSenti-SemEval shared task, where we tackle the task of sentiment analysis in 14 different African languages. We develop both monolingual and multilingual models under a full supervised setting (subtasks A and B). We also develop models for the zero-shot setting (subtask C). Our approach involves experimenting with transfer learning using six language models, including further pertaining of some of these models as well as a final finetuning stage. Our best performing models achieve an F1-score of 70.36 on development data and an F1-score of 66.13 on test data. Unsurprisingly, our results demonstrate the effectiveness of transfer learning and fine-tuning techniques for sentiment analysis across multiple languages. Our approach can be applied to other sentiment analysis tasks in different languages and domains.

Gagan Bhatia, El Moatez Billah Nagoudi, Hasan Cavusoglu et al.

We introduce FinTral, a suite of state-of-the-art multimodal large language models (LLMs) built upon the Mistral-7b model and tailored for financial analysis. FinTral integrates textual, numerical, tabular, and image data. We enhance FinTral with domain-specific pretraining, instruction fine-tuning, and RLAIF training by exploiting a large collection of textual and visual datasets we curate for this work. We also introduce an extensive benchmark featuring nine tasks and 25 datasets for evaluation, including hallucinations in the financial domain. Our FinTral model trained with direct preference optimization employing advanced Tools and Retrieval methods, dubbed FinTral-DPO-T&R, demonstrates an exceptional zero-shot performance. It outperforms ChatGPT-3.5 in all tasks and surpasses GPT-4 in five out of nine tasks, marking a significant advancement in AI-driven financial technology. We also demonstrate that FinTral has the potential to excel in real-time analysis and decision-making in diverse financial contexts. The GitHub repository for FinTral is available at \url{https://github.com/UBC-NLP/fintral}.

Fakhraddin Alwajih, El Moatez Billah Nagoudi, Gagan Bhatia et al.

Multimodal large language models (MLLMs) have proven effective in a wide range of tasks requiring complex reasoning and linguistic comprehension. However, due to a lack of high-quality multimodal resources in languages other than English, success of MLLMs remains relatively limited to English-based settings. This poses significant challenges in developing comparable models for other languages, including even those with large speaker populations such as Arabic. To alleviate this challenge, we introduce a comprehensive family of Arabic MLLMs, dubbed \textit{Peacock}, with strong vision and language capabilities. Through comprehensive qualitative and quantitative analysis, we demonstrate the solid performance of our models on various visual reasoning tasks and further show their emerging dialectal potential. Additionally, we introduce ~\textit{Henna}, a new benchmark specifically designed for assessing MLLMs on aspects related to Arabic culture, setting the first stone for culturally-aware Arabic MLLMs.The GitHub repository for the \textit{Peacock} project is available at \url{https://github.com/UBC-NLP/peacock}.

Gagan Bhatia, El Moatez Billah Nagoudi, Abdellah El Mekki et al.

We introduce {\bf Swan}, a family of embedding models centred around the Arabic language, addressing both small-scale and large-scale use cases. Swan includes two variants: Swan-Small, based on ARBERTv2, and Swan-Large, built on ArMistral, a pretrained Arabic large language model. To evaluate these models, we propose ArabicMTEB, a comprehensive benchmark suite that assesses cross-lingual, multi-dialectal, multi-domain, and multi-cultural Arabic text embedding performance, covering eight diverse tasks and spanning 94 datasets. Swan-Large achieves state-of-the-art results, outperforming Multilingual-E5-large in most Arabic tasks, while the Swan-Small consistently surpasses Multilingual-E5-base. Our extensive evaluations demonstrate that Swan models are both dialectally and culturally aware, excelling across various Arabic domains while offering significant monetary efficiency. This work significantly advances the field of Arabic language modelling and provides valuable resources for future research and applications in Arabic natural language processing. Our models and benchmark are available at our GitHub page: \href{https://github.com/UBC-NLP/swan}{https://github.com/UBC-NLP/swan}

Gagan Bhatia, Maxime Peyrard, Wei Zhao

Modern BPE tokenizers often split calendar dates into meaningless fragments, e.g., 20250312 $\rightarrow$ 202, 503, 12, inflating token counts and obscuring the inherent structure needed for robust temporal reasoning. In this work, we (1) introduce a simple yet interpretable metric, termed date fragmentation ratio, that measures how faithfully a tokenizer preserves multi-digit date components; (2) release DateAugBench, a suite of 6500 examples spanning three temporal reasoning tasks: context-based date resolution, format-invariance puzzles, and date arithmetic across historical, contemporary, and future time periods; and (3) through layer-wise probing and causal attention-hop analyses, uncover an emergent date-abstraction mechanism whereby large language models stitch together the fragments of month, day, and year components for temporal reasoning. Our experiments show that excessive fragmentation correlates with accuracy drops of up to 10 points on uncommon dates like historical and futuristic dates. Further, we find that the larger the model, the faster the emergent date abstraction that heals date fragments is accomplished. Lastly, we observe a reasoning path that LLMs follow to assemble date fragments, typically differing from human interpretation (year $\rightarrow$ month $\rightarrow$ day). Our datasets and code are made publicly available \href{https://github.com/gagan3012/date-fragments}{here}.

Subhadeep Roy, Gagan Bhatia, Steffen Eger

Automatic metrics are now central to evaluating text-to-image models, often substituting for human judgment in benchmarking and large-scale filtering. However, it remains unclear whether these metrics truly prioritize semantic correctness or instead favor visually and socially prototypical images learned from biased data distributions. We identify and study prototypicality bias as a systematic failure mode in multimodal evaluation. We introduce a controlled contrastive benchmark ProtoBias (Prototypical Bias), spanning Animals, Objects, and Demography images, where semantically correct but non-prototypical images are paired with subtly incorrect yet prototypical adversarial counterparts. This setup enables a directional evaluation of whether metrics follow textual semantics or default to prototypes. Our results show that widely used metrics, including CLIPScore, PickScore, and VQA-based scores, frequently misrank these pairs, while even LLM-as-Judge systems exhibit uneven robustness in socially grounded cases. Human evaluations consistently favour semantic correctness with larger decision margins. Motivated by these findings, we propose ProtoScore, a robust 7B-parameter metric that substantially reduces failure rates and suppresses misranking, while running at orders of magnitude faster than the inference time of GPT-5, approaching the robustness of much larger closed-source judges.

Sang Yun Kwon, Gagan Bhatia, El Moatez Billah Nagoudi et al.

Large language models (LLMs) finetuned to follow human instruction have recently exhibited significant capabilities in various English NLP tasks. However, their performance in grammatical error correction (GEC), especially on languages other than English, remains significantly unexplored. In this work, we evaluate the abilities of instruction finetuned LLMs in Arabic GEC, a complex task due to Arabic's rich morphology. Our findings suggest that various prompting methods, coupled with (in-context) few-shot learning, demonstrate considerable effectiveness, with GPT-4 achieving up to $65.49$ F$_{1}$ score under expert prompting (approximately $5$ points higher than our established baseline). Despite these positive results, we find that instruction finetuned models, regardless of their size, are still outperformed by fully finetuned ones, even if they are significantly smaller in size. This disparity highlights substantial room for improvements for LLMs. Inspired by methods used in low-resource machine translation, we also develop a method exploiting synthetic data that significantly outperforms previous models on two standard Arabic benchmarks. Our best model achieves a new SOTA on Arabic GEC, with $73.29$ and $73.26$ F$_{1}$ on the 2014 and 2015 QALB datasets, respectively, compared to peer-reviewed published baselines.

Gagan Bhatia, MingZe Tang, Cristina Mahanta et al.

This paper introduces DateLogicQA, a benchmark with 190 questions covering diverse date formats, temporal contexts, and reasoning types. We propose the Semantic Integrity Metric to assess tokenization quality and analyse two biases: Representation-Level Bias, affecting embeddings, and Logical-Level Bias, influencing reasoning outputs. Our findings provide a comprehensive evaluation of LLMs' capabilities and limitations in temporal reasoning, highlighting key challenges in handling temporal data accurately.

Gagan Bhatia, Somayajulu G Sripada, Kevin Allan et al.

Large Language Models (LLMs) are prone to hallucination, the generation of plausible yet factually incorrect statements. This work investigates the intrinsic, architectural origins of this failure mode through three primary contributions. First, to enable the reliable tracing of internal semantic failures, we propose Distributional Semantics Tracing (DST), a unified framework that integrates established interpretability techniques to produce a causal map of a model's reasoning, treating meaning as a function of context (distributional semantics). Second, we pinpoint the model's layer at which a hallucination becomes inevitable, identifying a specific commitment layer where a model's internal representations irreversibly diverge from factuality. Third, we identify the underlying mechanism for these failures. We observe a conflict between distinct computational pathways, which we interpret using the lens of dual-process theory: a fast, heuristic associative pathway (akin to System 1) and a slow, deliberate, contextual pathway (akin to System 2), leading to predictable failure modes such as Reasoning Shortcut Hijacks. Our framework's ability to quantify the coherence of the contextual pathway reveals a strong negative correlation ($ρ= -0.863$) with hallucination rates, implying that these failures are predictable consequences of internal semantic weakness. The result is a mechanistic account of how, when, and why hallucinations occur within the Transformer architecture.

Cristina Mahanta, Gagan Bhatia

Recognising human activity in a single photo enables indexing, safety and assistive applications, yet lacks motion cues. Using 285 MSCOCO images labelled as walking, running, sitting, and standing, scratch CNNs scored 41% accuracy. Fine-tuning multimodal CLIP raised this to 76%, demonstrating that contrastive vision-language pre-training decisively improves still-image action recognition in real-world deployments.