Ratish Puduppully

Nandini Mundra, Aditya Nanda Kishore, Raj Dabre et al. · microsoft-research

Language Models (LMs) excel in natural language processing tasks for English but show reduced performance in most other languages. This problem is commonly tackled by continually pre-training and fine-tuning these models for said languages. A significant issue in this process is the limited vocabulary coverage in the original model's tokenizer, leading to inadequate representation of new languages and necessitating an expansion of the tokenizer. The initialization of the embeddings corresponding to new vocabulary items presents a further challenge. Current strategies require cross-lingual embeddings and lack a solid theoretical foundation as well as comparisons with strong baselines. In this paper, we first establish theoretically that initializing within the convex hull of existing embeddings is a good initialization, followed by a novel but simple approach, Constrained Word2Vec (CW2V), which does not require cross-lingual embeddings. Our study evaluates different initialization methods for expanding RoBERTa and LLaMA 2 across four languages and five tasks. The results show that CW2V performs equally well or even better than more advanced techniques. Additionally, simpler approaches like multivariate initialization perform on par with these advanced methods indicating that efficient large-scale multilingual continued pretraining can be achieved even with simpler initialization methods. We release our code publicly (https://github.com/AI4Bharat/VocabAdaptation_LLM/tree/CW2V).

Sebastian Gehrmann, Abhik Bhattacharjee, Abinaya Mahendiran et al. · amazon-science, cmu

Evaluation in machine learning is usually informed by past choices, for example which datasets or metrics to use. This standardization enables the comparison on equal footing using leaderboards, but the evaluation choices become sub-optimal as better alternatives arise. This problem is especially pertinent in natural language generation which requires ever-improving suites of datasets, metrics, and human evaluation to make definitive claims. To make following best model evaluation practices easier, we introduce GEMv2. The new version of the Generation, Evaluation, and Metrics Benchmark introduces a modular infrastructure for dataset, model, and metric developers to benefit from each others work. GEMv2 supports 40 documented datasets in 51 languages. Models for all datasets can be evaluated online and our interactive data card creation and rendering tools make it easier to add new datasets to the living benchmark.

Thanmay Jayakumar, Mohammed Safi Ur Rahman Khan, Raj Dabre et al. · microsoft-research

Instruction-following benchmarks remain predominantly English-centric, leaving a critical evaluation gap for the hundreds of millions of Indic language speakers. We introduce IndicIFEval, a benchmark evaluating constrained generation of LLMs across 14 Indic languages using automatically verifiable, rule-based instructions. It comprises around 800 human-verified examples per language spread across two complementary subsets: IndicIFEval-Ground, translated prompts from IFEval (Zhou et al., 2023) carefully localized for Indic contexts, and IndicIFEval-Ground, synthetically generated instructions grounded in native Indic content. We conduct a comprehensive evaluation of major open-weight and proprietary models spanning both reasoning and non-reasoning models. While models maintain strong adherence to formatting constraints, they struggle significantly with lexical and cross-lingual tasks -- and despite progress in high-resource languages, instruction-following across the broader Indic family lags significantly behind English. We release IndicIFEval and its evaluation scripts to support progress on multilingual constrained generation (http://github.com/ai4bharat/IndicIFEval).

Aman Kumar, Himani Shrotriya, Prachi Sahu et al. · microsoft-research

Natural Language Generation (NLG) for non-English languages is hampered by the scarcity of datasets in these languages. In this paper, we present the IndicNLG Benchmark, a collection of datasets for benchmarking NLG for 11 Indic languages. We focus on five diverse tasks, namely, biography generation using Wikipedia infoboxes, news headline generation, sentence summarization, paraphrase generation and, question generation. We describe the created datasets and use them to benchmark the performance of several monolingual and multilingual baselines that leverage pre-trained sequence-to-sequence models. Our results exhibit the strong performance of multilingual language-specific pre-trained models, and the utility of models trained on our dataset for other related NLG tasks. Our dataset creation methods can be easily applied to modest-resource languages as they involve simple steps such as scraping news articles and Wikipedia infoboxes, light cleaning, and pivoting through machine translation data. To the best of our knowledge, the IndicNLG Benchmark is the first NLG benchmark for Indic languages and the most diverse multilingual NLG dataset, with approximately 8M examples across 5 tasks and 11 languages. The datasets and models are publicly available at https://ai4bharat.iitm.ac.in/indicnlg-suite.

Ratish Puduppully, Parag Jain, Nancy F. Chen et al.

In Multi-Document Summarization (MDS), the input can be modeled as a set of documents, and the output is its summary. In this paper, we focus on pretraining objectives for MDS. Specifically, we introduce a novel pretraining objective, which involves selecting the ROUGE-based centroid of each document cluster as a proxy for its summary. Our objective thus does not require human written summaries and can be utilized for pretraining on a dataset consisting solely of document sets. Through zero-shot, few-shot, and fully supervised experiments on multiple MDS datasets, we show that our model Centrum is better or comparable to a state-of-the-art model. We make the pretrained and fine-tuned models freely available to the research community https://github.com/ratishsp/centrum.

Vernon Toh Yan Han, Ratish Puduppully, Nancy F. Chen

Large Language Models (LLMs), combined with program-based solving techniques, are increasingly demonstrating proficiency in mathematical reasoning. For example, closed-source models such as OpenAI GPT-4 and Claude show excellent results in solving math word problems. However, progress in math word problem-solving for open-source LLMs is limited, and the challenges these models face are not well-studied. In this paper, we study the performance of strong open-source LLMs, including Llama 2 (7B), Code Llama (7B), and Mistral (7B) on math word problems using program-based solving techniques. Specifically, we analyze the outputs of these models when applied to math word problems and identify a category of problems that pose a significant challenge, particularly those involving quantities spanning multiple units. To address this issue, we propose a systematic approach by defining the units for each quantity and ensuring the consistency of these units during mathematical operations. We developed Unit Consistency Programs (UCPs), an annotated dataset of math word problems, each paired with programs containing unit specifications and unit verification routines. We fine-tuned Llama 2 (7B), Code Llama (7B), and Mistral (7B) models with UCPs to produce theirVerityMath variants. Our findings indicate that our approach, which incorporates unit consistency, currently slightly underperforms compared to an approach that does not. To understand the reasons behind this, we conduct an in-depth error analysis and suggest options for future improvements. Our code and dataset are available at https://github.com/vernontoh/VerityMath.

Jaavid Aktar Husain, Raj Dabre, Aswanth Kumar et al.

This study addresses the challenge of extending Large Language Models (LLMs) to non-English languages that use non-Roman scripts. We propose an approach that utilizes the romanized form of text as an interface for LLMs, hypothesizing that its frequent informal use and shared tokens with English enhance cross-lingual alignment. Our approach involves the continual pretraining of an English LLM like Llama 2 on romanized text of non-English, non-Roman script languages, followed by instruction tuning on romanized data. The results indicate that romanized text not only reduces token fertility by 2x-4x but also matches or outperforms native script representation across various NLU, NLG, and MT tasks. Moreover, the embeddings computed on romanized text exhibit closer alignment with their English translations than those from the native script. Our approach presents a promising direction for leveraging the power of English LLMs in languages traditionally underrepresented in NLP. Our code is available on https://github.com/AI4Bharat/romansetu.

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.

Nandini Mundra, Sumanth Doddapaneni, Raj Dabre et al.

Adapters have been positioned as a parameter-efficient fine-tuning (PEFT) approach, whereby a minimal number of parameters are added to the model and fine-tuned. However, adapters have not been sufficiently analyzed to understand if PEFT translates to benefits in training/deployment efficiency and maintainability/extensibility. Through extensive experiments on many adapters, tasks, and languages in supervised and cross-lingual zero-shot settings, we clearly show that for Natural Language Understanding (NLU) tasks, the parameter efficiency in adapters does not translate to efficiency gains compared to full fine-tuning of models. More precisely, adapters are relatively expensive to train and have slightly higher deployment latency. Furthermore, the maintainability/extensibility benefits of adapters can be achieved with simpler approaches like multi-task training via full fine-tuning, which also provide relatively faster training times. We, therefore, recommend that for moderately sized models for NLU tasks, practitioners should rely on full fine-tuning or multi-task training rather than using adapters. Our code is available at https://github.com/AI4Bharat/adapter-efficiency.

Alan Saji, Jaavid Aktar Husain, Thanmay Jayakumar et al. · microsoft-research

Large Language Models (LLMs) exhibit strong multilingual performance despite being predominantly trained on English-centric corpora. This raises a fundamental question: How do LLMs achieve such multilingual capabilities? Focusing on languages written in non-Roman scripts, we investigate the role of Romanization - the representation of non-Roman scripts using Roman characters - as a potential bridge in multilingual processing. Using mechanistic interpretability techniques, we analyze next-token generation and find that intermediate layers frequently represent target words in Romanized form before transitioning to native script, a phenomenon we term Latent Romanization. Further, through activation patching experiments, we demonstrate that LLMs encode semantic concepts similarly across native and Romanized scripts, suggesting a shared underlying representation. Additionally, for translation into non-Roman script languages, our findings reveal that when the target language is in Romanized form, its representations emerge earlier in the model's layers compared to native script. These insights contribute to a deeper understanding of multilingual representation in LLMs and highlight the implicit role of Romanization in facilitating language transfer.

Aditya Gorla, Ratish Puduppully

Tabular Language Models (TLMs) have been claimed to achieve emergent generalization for tabular prediction. We conduct a systematic re-evaluation of Tabula-8B as a representative TLM, utilizing 165 datasets from the UniPredict benchmark. Our investigation reveals three findings. First, binary and categorical classification achieve near-zero median lift over majority-class baselines and strong aggregate performance is driven entirely by quartile classification tasks. Second, top-performing datasets exhibit pervasive contamination, including complete train-test overlap and task-level leakage that evades standard deduplication. Third, instruction-tuning without tabular exposure recovers 92.2% of standard classification performance and on quartile classification, format familiarity closes 71.3% of the gap with the residual attributable to contaminated datasets. These findings suggest claimed generalization likely reflects evaluation artifacts rather than learned tabular reasoning. We conclude with recommendations for strengthening TLM evaluation.

Deepon Halder, Alan Saji, Thanmay Jayakumar et al. · microsoft-research

Large Language Models have demonstrated strong performance on many established reasoning benchmarks. However, these benchmarks primarily evaluate structured skills like quantitative problem-solving, leaving a gap in assessing flexible, multifaceted reasoning abilities that are central to human intelligence. These abilities require integrating logical deduction with spatial awareness and constraint satisfaction, which current evaluations do not measure well. To address this, we introduce RiddleBench, a benchmark of 1,737 challenging puzzles in English designed to probe these core reasoning capabilities. Evaluation of state-of-the-art models on RiddleBench shows fundamental weaknesses. Even top proprietary models like Gemini 2.5 Pro, o3, and Claude 4 Sonnet achieve accuracy just above 60% (60.30%, 63.37%, and 63.16%). Analysis further reveals deep failures, including hallucination cascades (accepting flawed reasoning from other models) and poor self-correction due to a strong self-confirmation bias. Their reasoning is also fragile, with performance degrading significantly when constraints are reordered or irrelevant information is introduced. RiddleBench functions as a diagnostic tool for these issues and as a resource for guiding the development of more robust and reliable language models.

Alan Saji, Raj Dabre, Anoop Kunchukuttan et al. · microsoft-research

Large Reasoning Models (LRMs) achieve strong performance on mathematical, scientific, and other question-answering tasks, but their multilingual reasoning abilities remain underexplored. When presented with non-English questions, LRMs often default to reasoning in English, raising concerns about interpretability and the handling of linguistic and cultural nuances. We systematically compare an LRM's reasoning in English versus the language of the question. Our evaluation spans two tasks: MGSM and GPQA Diamond. Beyond measuring answer accuracy, we also analyze cognitive attributes in the reasoning traces. We find that English reasoning traces exhibit a substantially higher presence of these cognitive behaviors, and that reasoning in English generally yields higher final-answer accuracy, with the performance gap increasing as tasks become more complex. However, this English-centric strategy is susceptible to a key failure mode - getting "Lost in Translation," where translation steps lead to errors that would have been avoided by question's language reasoning.

Aakash Lahoti, Tanya Marwah, Ratish Puduppully et al.

Transformer-based deep learning methods have become the standard approach for modeling diverse data such as sequences, images, and graphs. These methods rely on self-attention, which treats data as an unordered set of elements. This ignores the neighborhood structure or graph topology of the data and requires inductive biases--such as position embeddings in sequences and images, or random walks in graphs--to incorporate topology. However, designing such task-specific biases requires significant effort and can introduce side effects that hinder generalization. We introduce Chimera, a unified model that directly incorporates data topology in a principled way, removing the need for domain-specific biases. The key idea is that state space models--which naturally do not require position embeddings--can be generalized to capture any graph topology. Our experiments show that Chimera achieves strong performance across language, vision, and graph domains, outperforming BERT on GLUE by 0.7 points, ViT on ImageNet-1k by 2.6%, and all baselines on the Long Range Graph Benchmark. We further propose algorithmic optimizations to improve Chimera's efficiency: (1) for Directed Acyclic Graphs, Chimera can be implemented as a linear-time recurrence; (2) for general graphs, a simple mathematical relaxation achieves Transformer's quadratic complexity without domain-specific heuristics. These results validate Chimera's core contribution and support the idea that data topology is a powerful inductive bias across modalities.

Anushka Singh, Ananya B. Sai, Raj Dabre et al.

While machine translation evaluation has been studied primarily for high-resource languages, there has been a recent interest in evaluation for low-resource languages due to the increasing availability of data and models. In this paper, we focus on a zero-shot evaluation setting focusing on low-resource Indian languages, namely Assamese, Kannada, Maithili, and Punjabi. We collect sufficient Multi-Dimensional Quality Metrics (MQM) and Direct Assessment (DA) annotations to create test sets and meta-evaluate a plethora of automatic evaluation metrics. We observe that even for learned metrics, which are known to exhibit zero-shot performance, the Kendall Tau and Pearson correlations with human annotations are only as high as 0.32 and 0.45. Synthetic data approaches show mixed results and overall do not help close the gap by much for these languages. This indicates that there is still a long way to go for low-resource evaluation.

Jay Gala, Thanmay Jayakumar, Jaavid Aktar Husain et al.

We announce the initial release of "Airavata," an instruction-tuned LLM for Hindi. Airavata was created by fine-tuning OpenHathi with diverse, instruction-tuning Hindi datasets to make it better suited for assistive tasks. Along with the model, we also share the IndicInstruct dataset, which is a collection of diverse instruction-tuning datasets to enable further research for Indic LLMs. Additionally, we present evaluation benchmarks and a framework for assessing LLM performance across tasks in Hindi. Currently, Airavata supports Hindi, but we plan to expand this to all 22 scheduled Indic languages. You can access all artifacts at https://ai4bharat.github.io/airavata.

Aswanth Kumar, Ratish Puduppully, Raj Dabre et al.

Large language models have demonstrated the capability to perform on machine translation when the input is prompted with a few examples (in-context learning). Translation quality depends on various features of the selected examples, such as their quality and relevance, but previous work has predominantly focused on individual features in isolation. In this paper, we propose a general framework for combining different features influencing example selection. We learn a regression model, CTQ Scorer (Contextual Translation Quality), that selects examples based on multiple features in order to maximize the translation quality. On multiple language pairs and language models, we show that CTQ Scorer helps significantly outperform random selection as well as strong single-factor baselines reported in the literature. We also see an improvement of over 2.5 COMET points on average with respect to a strong BM25 retrieval-based baseline.

Ratish Puduppully, Anoop Kunchukuttan, Raj Dabre et al.

This study investigates machine translation between related languages i.e., languages within the same family that share linguistic characteristics such as word order and lexical similarity. Machine translation through few-shot prompting leverages a small set of translation pair examples to generate translations for test sentences. This procedure requires the model to learn how to generate translations while simultaneously ensuring that token ordering is maintained to produce a fluent and accurate translation. We propose that for related languages, the task of machine translation can be simplified by leveraging the monotonic alignment characteristic of such languages. We introduce DecoMT, a novel approach of few-shot prompting that decomposes the translation process into a sequence of word chunk translations. Through automatic and human evaluation conducted on multiple related language pairs across various language families, we demonstrate that our proposed approach of decomposed prompting surpasses multiple established few-shot baseline approaches. For example, DecoMT outperforms the strong few-shot prompting BLOOM model with an average improvement of 8 chrF++ scores across the examined languages.

Ratish Puduppully, Yao Fu, Mirella Lapata

We consider the task of data-to-text generation, which aims to create textual output from non-linguistic input. We focus on generating long-form text, i.e., documents with multiple paragraphs, and propose a neural model enhanced with a planning component responsible for organizing high-level information in a coherent and meaningful way. We infer latent plans sequentially with a structured variational model, while interleaving the steps of planning and generation. Text is generated by conditioning on previous variational decisions and previously generated text. Experiments on two data-to-text benchmarks (RotoWire and MLB) show that our model outperforms strong baselines and is sample efficient in the face of limited training data (e.g., a few hundred instances).

Raj Dabre, Himani Shrotriya, Anoop Kunchukuttan et al.

In this paper, we study pre-trained sequence-to-sequence models for a group of related languages, with a focus on Indic languages. We present IndicBART, a multilingual, sequence-to-sequence pre-trained model focusing on 11 Indic languages and English. IndicBART utilizes the orthographic similarity between Indic scripts to improve transfer learning between similar Indic languages. We evaluate IndicBART on two NLG tasks: Neural Machine Translation (NMT) and extreme summarization. Our experiments on NMT and extreme summarization show that a model specific to related languages like IndicBART is competitive with large pre-trained models like mBART50 despite being significantly smaller. It also performs well on very low-resource translation scenarios where languages are not included in pre-training or fine-tuning. Script sharing, multilingual training, and better utilization of limited model capacity contribute to the good performance of the compact IndicBART model.

Ratish Puduppully, Mirella Lapata

Recent approaches to data-to-text generation have adopted the very successful encoder-decoder architecture or variants thereof. These models generate text which is fluent (but often imprecise) and perform quite poorly at selecting appropriate content and ordering it coherently. To overcome some of these issues, we propose a neural model with a macro planning stage followed by a generation stage reminiscent of traditional methods which embrace separate modules for planning and surface realization. Macro plans represent high level organization of important content such as entities, events and their interactions; they are learnt from data and given as input to the generator. Extensive experiments on two data-to-text benchmarks (RotoWire and MLB) show that our approach outperforms competitive baselines in terms of automatic and human evaluation.

Ratish Puduppully, Yue Zhang, Manish Shrivastava

Traditional methods for deep NLG adopt pipeline approaches comprising stages such as constructing syntactic input, predicting function words, linearizing the syntactic input and generating the surface forms. Though easier to visualize, pipeline approaches suffer from error propagation. In addition, information available across modules cannot be leveraged by all modules. We construct a transition-based model to jointly perform linearization, function word prediction and morphological generation, which considerably improves upon the accuracy compared to a pipelined baseline system. On a standard deep input linearization shared task, our system achieves the best results reported so far.

Ratish Puduppully, Li Dong, Mirella Lapata

Recent approaches to data-to-text generation have shown great promise thanks to the use of large-scale datasets and the application of neural network architectures which are trained end-to-end. These models rely on representation learning to select content appropriately, structure it coherently, and verbalize it grammatically, treating entities as nothing more than vocabulary tokens. In this work we propose an entity-centric neural architecture for data-to-text generation. Our model creates entity-specific representations which are dynamically updated. Text is generated conditioned on the data input and entity memory representations using hierarchical attention at each time step. We present experiments on the RotoWire benchmark and a (five times larger) new dataset on the baseball domain which we create. Our results show that the proposed model outperforms competitive baselines in automatic and human evaluation.

Ratish Puduppully, Li Dong, Mirella Lapata

Recent advances in data-to-text generation have led to the use of large-scale datasets and neural network models which are trained end-to-end, without explicitly modeling what to say and in what order. In this work, we present a neural network architecture which incorporates content selection and planning without sacrificing end-to-end training. We decompose the generation task into two stages. Given a corpus of data records (paired with descriptive documents), we first generate a content plan highlighting which information should be mentioned and in which order and then generate the document while taking the content plan into account. Automatic and human-based evaluation experiments show that our model outperforms strong baselines improving the state-of-the-art on the recently released RotoWire dataset.