Varun Gumma

Rishav Hada, Varun Gumma, Adrian de Wynter et al. · microsoft-research

Large Language Models (LLMs) excel in various Natural Language Processing (NLP) tasks, yet their evaluation, particularly in languages beyond the top $20$, remains inadequate due to existing benchmarks and metrics limitations. Employing LLMs as evaluators to rank or score other models' outputs emerges as a viable solution, addressing the constraints tied to human annotators and established benchmarks. In this study, we explore the potential of LLM-based evaluators, specifically GPT-4 in enhancing multilingual evaluation by calibrating them against $20$K human judgments across three text-generation tasks, five metrics, and eight languages. Our analysis reveals a bias in GPT4-based evaluators towards higher scores, underscoring the necessity of calibration with native speaker judgments, especially in low-resource and non-Latin script languages, to ensure accurate evaluation of LLM performance across diverse languages.

Sanchit Ahuja, Divyanshu Aggarwal, Varun Gumma et al. · cmu, deepmind

There has been a surge in LLM evaluation research to understand LLM capabilities and limitations. However, much of this research has been confined to English, leaving LLM building and evaluation for non-English languages relatively unexplored. Several new LLMs have been introduced recently, necessitating their evaluation on non-English languages. This study aims to perform a thorough evaluation of the non-English capabilities of SoTA LLMs (GPT-3.5-Turbo, GPT-4, PaLM2, Gemini-Pro, Mistral, Llama2, and Gemma) by comparing them on the same set of multilingual datasets. Our benchmark comprises 22 datasets covering 83 languages, including low-resource African languages. We also include two multimodal datasets in the benchmark and compare the performance of LLaVA models, GPT-4-Vision and Gemini-Pro-Vision. Our experiments show that larger models such as GPT-4, Gemini-Pro and PaLM2 outperform smaller models on various tasks, notably on low-resource languages, with GPT-4 outperforming PaLM2 and Gemini-Pro on more datasets. We also perform a study on data contamination and find that several models are likely to be contaminated with multilingual evaluation benchmarks, necessitating approaches to detect and handle contamination while assessing the multilingual performance of LLMs.

Varun Gumma, Raj Dabre, Pratyush Kumar · microsoft-research

Knowledge distillation (KD) is a well-known method for compressing neural models. However, works focusing on distilling knowledge from large multilingual neural machine translation (MNMT) models into smaller ones are practically nonexistent, despite the popularity and superiority of MNMT. This paper bridges this gap by presenting an empirical investigation of knowledge distillation for compressing MNMT models. We take Indic to English translation as a case study and demonstrate that commonly used language-agnostic and language-aware KD approaches yield models that are 4-5x smaller but also suffer from performance drops of up to 3.5 BLEU. To mitigate this, we then experiment with design considerations such as shallower versus deeper models, heavy parameter sharing, multi-stage training, and adapters. We observe that deeper compact models tend to be as good as shallower non-compact ones, and that fine-tuning a distilled model on a High-Quality subset slightly boosts translation quality. Overall, we conclude that compressing MNMT models via KD is challenging, indicating immense scope for further research.

Varun Gumma, Pranjal A. Chitale, Kalika Bali · microsoft-research

Neural Machine Translation (NMT) models have traditionally used Sinusoidal Positional Embeddings (PEs), which often struggle to capture long-range dependencies and are inefficient for handling extended context or document-level translation tasks. This work addresses the challenge of transitioning pre-trained NMT models from absolute Sinusoidal PEs to Relative PEs, such as RoPE and ALiBi, without compromising performance. We demonstrate that parameter-efficient fine-tuning, using only a small amount of high-quality data, can successfully facilitate this transition. Experimental results indicate that switching from Sinusoidal to Relative PEs results in competitive translation quality on sentence-level evaluation benchmarks. Additionally, models trained with RoPE consistently outperform those using ALiBi and Sinusoidal PEs on document-level benchmarks across both string-based metrics and qualitative evaluations. Moreover, we find that a small amount of long-context data in a few languages is sufficient for cross-lingual length generalization, thereby inducing long-context capabilities.

Pranjal A. Chitale, Varun Gumma, Sanchit Ahuja et al. · microsoft-research

Developing AI systems that operate effectively across languages while remaining culturally grounded is a long-standing challenge, particularly in low-resource settings. Synthetic data provides a promising avenue, yet its effectiveness in multilingual and multicultural contexts remains underexplored. We investigate the creation and impact of synthetic, culturally contextualized datasets for Indian languages through a bottom-up generation strategy that prompts large open-source LLMs (>= 235B parameters) to ground data generation in language-specific Wikipedia content. This approach complements the dominant top-down paradigm of translating synthetic datasets from high-resource languages such as English. We introduce Updesh, a high-quality large-scale synthetic instruction-following dataset comprising 9.5M data points across 13 Indian languages, encompassing diverse reasoning and generative tasks with an emphasis on long-context, multi-turn capabilities, and alignment with Indian cultural contexts. A comprehensive evaluation incorporating both automated metrics and human annotation across 10k assessments indicates that generated data is high quality; though, human evaluation highlights areas for further improvement. Additionally, we perform downstream evaluations by fine-tuning models on our dataset and assessing the performance across 15 diverse multilingual datasets. Models trained on Updesh consistently achieve significant gains on generative tasks and remain competitive on multiple-choice style NLU tasks. Notably, relative improvements are most pronounced in low and medium-resource languages, narrowing their gap with high-resource languages. These findings provide empirical evidence that effective multilingual AI requires multi-faceted data curation and generation strategies that incorporate context-aware, culturally grounded methodologies.

Jay Gala, Pranjal A. Chitale, Raghavan AK et al.

India has a rich linguistic landscape with languages from 4 major language families spoken by over a billion people. 22 of these languages are listed in the Constitution of India (referred to as scheduled languages) are the focus of this work. Given the linguistic diversity, high-quality and accessible Machine Translation (MT) systems are essential in a country like India. Prior to this work, there was (i) no parallel training data spanning all 22 languages, (ii) no robust benchmarks covering all these languages and containing content relevant to India, and (iii) no existing translation models which support all the 22 scheduled languages of India. In this work, we aim to address this gap by focusing on the missing pieces required for enabling wide, easy, and open access to good machine translation systems for all 22 scheduled Indian languages. We identify four key areas of improvement: curating and creating larger training datasets, creating diverse and high-quality benchmarks, training multilingual models, and releasing models with open access. Our first contribution is the release of the Bharat Parallel Corpus Collection (BPCC), the largest publicly available parallel corpora for Indic languages. BPCC contains a total of 230M bitext pairs, of which a total of 126M were newly added, including 644K manually translated sentence pairs created as part of this work. Our second contribution is the release of the first n-way parallel benchmark covering all 22 Indian languages, featuring diverse domains, Indian-origin content, and source-original test sets. Next, we present IndicTrans2, the first model to support all 22 languages, surpassing existing models on multiple existing and new benchmarks created as a part of this work. Lastly, to promote accessibility and collaboration, we release our models and associated data with permissive licenses at https://github.com/AI4Bharat/IndicTrans2.

Rishav Hada, Safiya Husain, Varun Gumma et al. · microsoft-research

Existing research in measuring and mitigating gender bias predominantly centers on English, overlooking the intricate challenges posed by non-English languages and the Global South. This paper presents the first comprehensive study delving into the nuanced landscape of gender bias in Hindi, the third most spoken language globally. Our study employs diverse mining techniques, computational models, field studies and sheds light on the limitations of current methodologies. Given the challenges faced with mining gender biased statements in Hindi using existing methods, we conducted field studies to bootstrap the collection of such sentences. Through field studies involving rural and low-income community women, we uncover diverse perceptions of gender bias, underscoring the necessity for context-specific approaches. This paper advocates for a community-centric research design, amplifying voices often marginalized in previous studies. Our findings not only contribute to the understanding of gender bias in Hindi but also establish a foundation for further exploration of Indic languages. By exploring the intricacies of this understudied context, we call for thoughtful engagement with gender bias, promoting inclusivity and equity in linguistic and cultural contexts beyond the Global North.

Rishav Hada, Varun Gumma, Mohamed Ahmed et al. · microsoft-research

With the rising human-like precision of Large Language Models (LLMs) in numerous tasks, their utilization in a variety of real-world applications is becoming more prevalent. Several studies have shown that LLMs excel on many standard NLP benchmarks. However, it is challenging to evaluate LLMs due to test dataset contamination and the limitations of traditional metrics. Since human evaluations are difficult to collect, there is a growing interest in the community to use LLMs themselves as reference-free evaluators for subjective metrics. However, past work has shown that LLM-based evaluators can exhibit bias and have poor alignment with human judgments. In this study, we propose a framework for an end-to-end assessment of LLMs as evaluators in multilingual scenarios. We create a carefully curated dataset, covering 10 languages containing native speaker judgments for the task of summarization. This dataset is created specifically to evaluate LLM-based evaluators, which we refer to as meta-evaluation (METAL). We compare the performance of LLM-based evaluators created using GPT-3.5-Turbo, GPT-4, and PaLM2. Our results indicate that LLM-based evaluators based on GPT-4 perform the best across languages, while GPT-3.5-Turbo performs poorly. Additionally, we perform an analysis of the reasoning provided by LLM-based evaluators and find that it often does not match the reasoning provided by human judges.

Varun Gumma, Ananditha Raghunath, Mohit Jain et al. · microsoft-research

Assessing the capabilities and limitations of large language models (LLMs) has garnered significant interest, yet the evaluation of multiple models in real-world scenarios remains rare. Multilingual evaluation often relies on translated benchmarks, which typically do not capture linguistic and cultural nuances present in the source language. This study provides an extensive assessment of 24 LLMs on real world data collected from Indian patients interacting with a medical chatbot in Indian English and 4 other Indic languages. We employ a uniform Retrieval Augmented Generation framework to generate responses, which are evaluated using both automated techniques and human evaluators on four specific metrics relevant to our application. We find that models vary significantly in their performance and that instruction tuned Indic models do not always perform well on Indic language queries. Further, we empirically show that factual correctness is generally lower for responses to Indic queries compared to English queries. Finally, our qualitative work shows that code-mixed and culturally relevant queries in our dataset pose challenges to evaluated models.

Sanchit Ahuja, Varun Gumma, Sunayana Sitaram · microsoft-research

Benchmark contamination refers to the presence of test datasets in Large Language Model (LLM) pre-training or post-training data. Contamination can lead to inflated scores on benchmarks, compromising evaluation results and making it difficult to determine the capabilities of models. In this work, we study the contamination of popular multilingual benchmarks in LLMs that support multiple languages. We use the Black Box test to determine whether $7$ frequently used multilingual benchmarks are contaminated in $7$ popular open and closed LLMs and find that almost all models show signs of being contaminated with almost all the benchmarks we test. Our findings can help the community determine the best set of benchmarks to use for multilingual evaluation.

Prachi Jain, Ashutosh Sathe, Varun Gumma et al. · microsoft-research

Pretrained Language Models (PLMs) are widely used in NLP for various tasks. Recent studies have identified various biases that such models exhibit and have proposed methods to correct these biases. However, most of the works address a limited set of bias dimensions independently such as gender, race, or religion. Moreover, the methods typically involve finetuning the full model to maintain the performance on the downstream task. In this work, we aim to modularly debias a pretrained language model across multiple dimensions. Previous works extensively explored debiasing PLMs using limited US-centric counterfactual data augmentation (CDA). We use structured knowledge and a large generative model to build a diverse CDA across multiple bias dimensions in a semi-automated way. We highlight how existing debiasing methods do not consider interactions between multiple societal biases and propose a debiasing model that exploits the synergy amongst various societal biases and enables multi-bias debiasing simultaneously. An extensive evaluation on multiple tasks and languages demonstrates the efficacy of our approach.

Varun Gumma, Rishav Hada, Aditya Yadavalli et al. · microsoft-research

We present MunTTS, an end-to-end text-to-speech (TTS) system specifically for Mundari, a low-resource Indian language of the Austo-Asiatic family. Our work addresses the gap in linguistic technology for underrepresented languages by collecting and processing data to build a speech synthesis system. We begin our study by gathering a substantial dataset of Mundari text and speech and train end-to-end speech models. We also delve into the methods used for training our models, ensuring they are efficient and effective despite the data constraints. We evaluate our system with native speakers and objective metrics, demonstrating its potential as a tool for preserving and promoting the Mundari language in the digital age.

Jingdi Lei, Varun Gumma, Rishabh Bhardwaj et al. · microsoft-research

Large Language Model (LLM) safety is one of the most pressing challenges for enabling wide-scale deployment. While most studies and global discussions focus on generic harms, such as models assisting users in harming themselves or others, enterprises face a more fundamental concern: whether LLM-based agents are safe for their intended use case. To address this, we introduce operational safety, defined as an LLM's ability to appropriately accept or refuse user queries when tasked with a specific purpose. We further propose OffTopicEval, an evaluation suite and benchmark for measuring operational safety both in general and within specific agentic use cases. Our evaluations on six model families comprising 20 open-weight LLMs reveal that while performance varies across models, all of them remain highly operationally unsafe. Even the strongest models - Qwen-3 (235B) with 77.77% and Mistral (24B) with 79.96% - fall far short of reliable operational safety, while GPT models plateau in the 62-73% range, Phi achieves only mid-level scores (48-70%), and Gemma and Llama-3 collapse to 39.53% and 23.84%, respectively. While operational safety is a core model alignment issue, to suppress these failures, we propose prompt-based steering methods: query grounding (Q-ground) and system-prompt grounding (P-ground), which substantially improve OOD refusal. Q-ground provides consistent gains of up to 23%, while P-ground delivers even larger boosts, raising Llama-3.3 (70B) by 41% and Qwen-3 (30B) by 27%. These results highlight both the urgent need for operational safety interventions and the promise of prompt-based steering as a first step toward more reliable LLM-based agents.

Ishaan Watts, Varun Gumma, Aditya Yadavalli et al.

Evaluation of multilingual Large Language Models (LLMs) is challenging due to a variety of factors -- the lack of benchmarks with sufficient linguistic diversity, contamination of popular benchmarks into LLM pre-training data and the lack of local, cultural nuances in translated benchmarks. In this work, we study human and LLM-based evaluation in a multilingual, multi-cultural setting. We evaluate 30 models across 10 Indic languages by conducting 90K human evaluations and 30K LLM-based evaluations and find that models such as GPT-4o and Llama-3 70B consistently perform best for most Indic languages. We build leaderboards for two evaluation settings - pairwise comparison and direct assessment and analyze the agreement between humans and LLMs. We find that humans and LLMs agree fairly well in the pairwise setting but the agreement drops for direct assessment evaluation especially for languages such as Bengali and Odia. We also check for various biases in human and LLM-based evaluation and find evidence of self-bias in the GPT-based evaluator. Our work presents a significant step towards scaling up multilingual evaluation of LLMs.

Millicent Ochieng, Varun Gumma, Sunayana Sitaram et al.

The deployment of Large Language Models (LLMs) in real-world applications presents both opportunities and challenges, particularly in multilingual and code-mixed communication settings. This research evaluates the performance of seven leading LLMs in sentiment analysis on a dataset derived from multilingual and code-mixed WhatsApp chats, including Swahili, English and Sheng. Our evaluation includes both quantitative analysis using metrics like F1 score and qualitative assessment of LLMs' explanations for their predictions. We find that, while Mistral-7b and Mixtral-8x7b achieved high F1 scores, they and other LLMs such as GPT-3.5-Turbo, Llama-2-70b, and Gemma-7b struggled with understanding linguistic and contextual nuances, as well as lack of transparency in their decision-making process as observed from their explanations. In contrast, GPT-4 and GPT-4-Turbo excelled in grasping diverse linguistic inputs and managing various contextual information, demonstrating high consistency with human alignment and transparency in their decision-making process. The LLMs however, encountered difficulties in incorporating cultural nuance especially in non-English settings with GPT-4s doing so inconsistently. The findings emphasize the necessity of continuous improvement of LLMs to effectively tackle the challenges of culturally nuanced, low-resource real-world settings and the need for developing evaluation benchmarks for capturing these issues.

Varun Gumma, Barsha Mitra, Soumyadeep Dey et al.

In recent years, Attribute-Based Access Control (ABAC) has become quite popular and effective for enforcing access control in dynamic and collaborative environments. Implementation of ABAC requires the creation of a set of attribute-based rules which cumulatively form a policy. Designing an ABAC policy ab initio demands a substantial amount of effort from the system administrator. Moreover, organizational changes may necessitate the inclusion of new rules in an already deployed policy. In such a case, re-mining the entire ABAC policy will require a considerable amount of time and administrative effort. Instead, it is better to incrementally augment the policy. Keeping these aspects of reducing administrative overhead in mind, in this paper, we propose PAMMELA, a Policy Administration Methodology using Machine Learning to help system administrators in creating new ABAC policies as well as augmenting existing ones. PAMMELA can generate a new policy for an organization by learning the rules of a policy currently enforced in a similar organization. For policy augmentation, PAMMELA can infer new rules based on the knowledge gathered from the existing rules. Experimental results show that our proposed approach provides a reasonably good performance in terms of the various machine learning evaluation metrics as well as execution time.