Ganesh Ramakrishnan

Renbo Tu, Nicholas Roberts, Vishak Prasad et al.

The challenge that climate change poses to humanity has spurred a rapidly developing field of artificial intelligence research focused on climate change applications. The climate change AI (CCAI) community works on a diverse, challenging set of problems which often involve physics-constrained ML or heterogeneous spatiotemporal data. It would be desirable to use automated machine learning (AutoML) techniques to automatically find high-performing architectures and hyperparameters for a given dataset. In this work, we benchmark popular AutoML libraries on three high-leverage CCAI applications: climate modeling, wind power forecasting, and catalyst discovery. We find that out-of-the-box AutoML libraries currently fail to meaningfully surpass the performance of human-designed CCAI models. However, we also identify a few key weaknesses, which stem from the fact that most AutoML techniques are tailored to computer vision and NLP applications. For example, while dozens of search spaces have been designed for image and language data, none have been designed for spatiotemporal data. Addressing these key weaknesses can lead to the discovery of novel architectures that yield substantial performance gains across numerous CCAI applications. Therefore, we present a call to action to the AutoML community, since there are a number of concrete, promising directions for future work in the space of AutoML for CCAI. We release our code and a list of resources at https://github.com/climate-change-automl/climate-change-automl.

Ayush Maheshwari, Nikhil Singh, Amrith Krishna et al.

Sanskrit is a classical language with about 30 million extant manuscripts fit for digitisation, available in written, printed or scannedimage forms. However, it is still considered to be a low-resource language when it comes to available digital resources. In this work, we release a post-OCR text correction dataset containing around 218,000 sentences, with 1.5 million words, from 30 different books. Texts in Sanskrit are known to be diverse in terms of their linguistic and stylistic usage since Sanskrit was the 'lingua franca' for discourse in the Indian subcontinent for about 3 millennia. Keeping this in mind, we release a multi-domain dataset, from areas as diverse as astronomy, medicine and mathematics, with some of them as old as 18 centuries. Further, we release multiple strong baselines as benchmarks for the task, based on pre-trained Seq2Seq language models. We find that our best-performing model, consisting of byte level tokenization in conjunction with phonetic encoding (Byt5+SLP1), yields a 23% point increase over the OCR output in terms of word and character error rates. Moreover, we perform extensive experiments in evaluating these models on their performance and analyse common causes of mispredictions both at the graphemic and lexical levels. Our code and dataset is publicly available at https://github.com/ayushbits/pe-ocr-sanskrit.

Ayush Maheshwari, Preethi Jyothi, Ganesh Ramakrishnan · deepmind

Domain-specific neural machine translation (NMT) systems (e.g., in educational applications) are socially significant with the potential to help make information accessible to a diverse set of users in multilingual societies. It is desirable that such NMT systems be lexically constrained and draw from domain-specific dictionaries. Dictionaries could present multiple candidate translations for a source word/phrase due to the polysemous nature of words. The onus is then on the NMT model to choose the contextually most appropriate candidate. Prior work has largely ignored this problem and focused on the single candidate constraint setting wherein the target word or phrase is replaced by a single constraint. In this work we present DictDis, a lexically constrained NMT system that disambiguates between multiple candidate translations derived from dictionaries. We achieve this by augmenting training data with multiple dictionary candidates to actively encourage disambiguation during training by implicitly aligning multiple candidate constraints. We demonstrate the utility of DictDis via extensive experiments on English-Hindi and English-German sentences in a variety of domains including regulatory, finance, engineering. We also present comparisons on standard benchmark test datasets. In comparison with existing approaches for lexically constrained and unconstrained NMT, we demonstrate superior performance with respect to constraint copy and disambiguation related measures on all domains while also obtaining improved fluency of up to 2-3 BLEU points on some domains.

Rishabh Tiwari, Durga Sivasubramanian, Anmol Mekala et al. · berkeley

Deep networks tend to learn spurious feature-label correlations in real-world supervised learning tasks. This vulnerability is aggravated in distillation, where a student model may have lesser representational capacity than the corresponding teacher model. Often, knowledge of specific spurious correlations is used to reweight instances & rebalance the learning process. We propose a novel early readout mechanism whereby we attempt to predict the label using representations from earlier network layers. We show that these early readouts automatically identify problem instances or groups in the form of confident, incorrect predictions. Leveraging these signals to modulate the distillation loss on an instance level allows us to substantially improve not only group fairness measures across benchmark datasets, but also overall accuracy of the student model. We also provide secondary analyses that bring insight into the role of feature learning in supervision and distillation.

Ayush Maheshwari, Ajay Ravindran, Venkatapathy Subramanian et al.

We introduce UDAAN, an open-source post-editing tool that can reduce manual editing efforts to quickly produce publishable-standard documents in several Indic languages. UDAAN has an end-to-end Machine Translation (MT) plus post-editing pipeline wherein users can upload a document to obtain raw MT output. Further, users can edit the raw translations using our tool. UDAAN offers several advantages: a) Domain-aware, vocabulary-based lexical constrained MT. b) source-target and target-target lexicon suggestions for users. Replacements are based on the source and target texts lexicon alignment. c) Translation suggestions are based on logs created during user interaction. d) Source-target sentence alignment visualisation that reduces the cognitive load of users during editing. e) Translated outputs from our tool are available in multiple formats: docs, latex, and PDF. We also provide the facility to use around 100 in-domain dictionaries for lexicon-aware machine translation. Although we limit our experiments to English-to-Hindi translation, our tool is independent of the source and target languages. Experimental results based on the usage of the tools and users feedback show that our tool speeds up the translation time by approximately a factor of three compared to the baseline method of translating documents from scratch. Our tool is available for both Windows and Linux platforms. The tool is open-source under MIT license, and the source code can be accessed from our website at https://www.udaanproject.org. Demonstration and tutorial videos for various features of our tool can be accessed at https://www.youtube.com/channel/UClfK7iC8J7b22bj3GwAUaCw. Our MT pipeline can be accessed at https://udaaniitb.aicte-india.org/udaan/translate/.

Prateek Chanda, Prayas Agrawal, Karthik S. Gurumoorthy et al.

Selecting prototypical examples from a source distribution to represent a target data distribution is a fundamental problem in machine learning. Existing subset selection methods often rely on implicit importance scores, which can be skewed towards majority classes and lead to low-quality prototypes for minority classes. We present $\methodprop$, a novel subset selection framework that minimizes the optimal transport (OT) distance between a uniformly weighted prototypical distribution and the target distribution. While intuitive, this formulation leads to a cardinality-constrained maximization of a \emph{super-additive} objective, which is generally intractable to approximate efficiently. To address this, we propose a principled reformulation of the OT marginal constraints, yielding a partial optimal transport-based submodular objective. We prove that this reformulation enables a greedy algorithm with a $(1-1/e)$ approximation guarantee relative to the original super-additive maximization problem. Empirically, we showcase that enforcing uniform prototype weights in UniPROT consistently improves minority-class representation in imbalanced classification benchmarks without compromising majority-class accuracy. In both finetuning and pretraining regimes for large language models under domain imbalance, UniPROT enforces uniform source contributions, yielding robust performance gains. Our results establish UniPROT as a scalable, theoretically grounded solution for uniform-weighted prototype selection. Our code is publicly available at GitHub\footnote{Code: https://github.com/efficiency-learning/UniPROT}

Krishnateja Killamsetty, Guttu Sai Abhishek, Aakriti et al.

Deep neural networks have seen great success in recent years; however, training a deep model is often challenging as its performance heavily depends on the hyper-parameters used. In addition, finding the optimal hyper-parameter configuration, even with state-of-the-art (SOTA) hyper-parameter optimization (HPO) algorithms, can be time-consuming, requiring multiple training runs over the entire dataset for different possible sets of hyper-parameters. Our central insight is that using an informative subset of the dataset for model training runs involved in hyper-parameter optimization, allows us to find the optimal hyper-parameter configuration significantly faster. In this work, we propose AUTOMATA, a gradient-based subset selection framework for hyper-parameter tuning. We empirically evaluate the effectiveness of AUTOMATA in hyper-parameter tuning through several experiments on real-world datasets in the text, vision, and tabular domains. Our experiments show that using gradient-based data subsets for hyper-parameter tuning achieves significantly faster turnaround times and speedups of 3$\times$-30$\times$ while achieving comparable performance to the hyper-parameters found using the entire dataset.

Mohd Nauman, Sravan Gvm, Vijay Devane et al.

Current large language models excel at broad, general-purpose tasks, but consistently underperform when exposed to highly specialized domains that require deep cultural, linguistic, and subject-matter expertise. In particular, traditional medical systems such as Ayurveda embody centuries of nuanced textual and clinical knowledge that mainstream LLMs fail to accurately interpret or apply. We introduce AyurParam-2.9B, a domain-specialized, bilingual language model fine-tuned from Param-1-2.9B using an extensive, expertly curated Ayurveda dataset spanning classical texts and clinical guidance. AyurParam's dataset incorporates context-aware, reasoning, and objective-style Q&A in both English and Hindi, with rigorous annotation protocols for factual precision and instructional clarity. Benchmarked on BhashaBench-Ayur, AyurParam not only surpasses all open-source instruction-tuned models in its size class (1.5--3B parameters), but also demonstrates competitive or superior performance compared to much larger models. The results from AyurParam highlight the necessity for authentic domain adaptation and high-quality supervision in delivering reliable, culturally congruent AI for specialized medical knowledge.

Abhishek Singh, Venkatapathy Subramanian, Ayush Maheshwari et al.

Information Extraction (IE) from document images is challenging due to the high variability of layout formats. Deep models such as LayoutLM and BROS have been proposed to address this problem and have shown promising results. However, they still require a large amount of field-level annotations for training these models. Other approaches using rule-based methods have also been proposed based on the understanding of the layout and semantics of a form such as geometric position, or type of the fields, etc. In this work, we propose a novel approach, EIGEN (Expert-Informed Joint Learning aGgrEatioN), which combines rule-based methods with deep learning models using data programming approaches to circumvent the requirement of annotation of large amounts of training data. Specifically, EIGEN consolidates weak labels induced from multiple heuristics through generative models and use them along with a small number of annotated labels to jointly train a deep model. In our framework, we propose the use of labeling functions that include incorporating contextual information thus capturing the visual and language context of a word for accurate categorization. We empirically show that our EIGEN framework can significantly improve the performance of state-of-the-art deep models with the availability of very few labeled data instances. The source code is available at https://github.com/ayushayush591/EIGEN-High-Fidelity-Extraction-Document-Images.

Suraj Kothawade, Atharv Savarkar, Venkat Iyer et al.

Training deep learning models on medical datasets that perform well for all classes is a challenging task. It is often the case that a suboptimal performance is obtained on some classes due to the natural class imbalance issue that comes with medical data. An effective way to tackle this problem is by using targeted active learning, where we iteratively add data points to the training data that belong to the rare classes. However, existing active learning methods are ineffective in targeting rare classes in medical datasets. In this work, we propose Clinical (targeted aCtive Learning for ImbalaNced medICal imAge cLassification) a framework that uses submodular mutual information functions as acquisition functions to mine critical data points from rare classes. We apply our framework to a wide-array of medical imaging datasets on a variety of real-world class imbalance scenarios - namely, binary imbalance and long-tail imbalance. We show that Clinical outperforms the state-of-the-art active learning methods by acquiring a diverse set of data points that belong to the rare classes.

Vishak Prasad C, Colin White, Paarth Jain et al.

A majority of recent developments in neural architecture search (NAS) have been aimed at decreasing the computational cost of various techniques without affecting their final performance. Towards this goal, several low-fidelity and performance prediction methods have been considered, including those that train only on subsets of the training data. In this work, we present an adaptive subset selection approach to NAS and present it as complementary to state-of-the-art NAS approaches. We uncover a natural connection between one-shot NAS algorithms and adaptive subset selection and devise an algorithm that makes use of state-of-the-art techniques from both areas. We use these techniques to substantially reduce the runtime of DARTS-PT (a leading one-shot NAS algorithm), as well as BOHB and DEHB (leading multifidelity optimization algorithms), without sacrificing accuracy. Our results are consistent across multiple datasets, and towards full reproducibility, we release our code at https: //anonymous.4open.science/r/SubsetSelection NAS-B132.

Ashish Mittal, Durga Sivasubramanian, Rishabh Iyer et al.

Training state-of-the-art ASR systems such as RNN-T often has a high associated financial and environmental cost. Training with a subset of training data could mitigate this problem if the subset selected could achieve on-par performance with training with the entire dataset. Although there are many data subset selection(DSS) algorithms, direct application to the RNN-T is difficult, especially the DSS algorithms that are adaptive and use learning dynamics such as gradients, as RNN-T tend to have gradients with a significantly larger memory footprint. In this paper, we propose Partitioned Gradient Matching (PGM) a novel distributable DSS algorithm, suitable for massive datasets like those used to train RNN-T. Through extensive experiments on Librispeech 100H and Librispeech 960H, we show that PGM achieves between 3x to 6x speedup with only a very small accuracy degradation (under 1% absolute WER difference). In addition, we demonstrate similar results for PGM even in settings where the training data is corrupted with noise.

Abhishek Kumar Singh, Vishwajeet kumar, Rudra Murthy et al.

Large Language Models (LLMs) perform well on unseen tasks in English, but their abilities in non English languages are less explored due to limited benchmarks and training data. To bridge this gap, we introduce the Indic QA Benchmark, a large dataset for context grounded question answering in 11 major Indian languages, covering both extractive and abstractive tasks. Evaluations of multilingual LLMs, including instruction finetuned versions, revealed weak performance in low resource languages due to a strong English language bias in their training data. We also investigated the Translate Test paradigm,where inputs are translated to English for processing and the results are translated back into the source language for output. This approach outperformed multilingual LLMs, particularly in low resource settings. By releasing Indic QA, we aim to promote further research into LLMs question answering capabilities in low resource languages. This benchmark offers a critical resource to address existing limitations and foster multilingual understanding.

Suraj Kothawade, Akshit Srivastava, Venkat Iyer et al.

Avoiding out-of-distribution (OOD) data is critical for training supervised machine learning models in the medical imaging domain. Furthermore, obtaining labeled medical data is difficult and expensive since it requires expert annotators like doctors, radiologists, etc. Active learning (AL) is a well-known method to mitigate labeling costs by selecting the most diverse or uncertain samples. However, current AL methods do not work well in the medical imaging domain with OOD data. We propose Diagnose (avoiDing out-of-dIstribution dAta usinG submodular iNfOrmation meaSurEs), a novel active learning framework that can jointly model similarity and dissimilarity, which is crucial in mining in-distribution data and avoiding OOD data at the same time. Particularly, we use a small number of data points as exemplars that represent a query set of in-distribution data points and a private set of OOD data points. We illustrate the generalizability of our framework by evaluating it on a wide variety of real-world OOD scenarios. Our experiments verify the superiority of Diagnose over the state-of-the-art AL methods across multiple domains of medical imaging.

Suraj Kothawade, Pavan Kumar Reddy, Ganesh Ramakrishnan et al.

Current semi-supervised learning (SSL) methods assume a balance between the number of data points available for each class in both the labeled and the unlabeled data sets. However, there naturally exists a class imbalance in most real-world datasets. It is known that training models on such imbalanced datasets leads to biased models, which in turn lead to biased predictions towards the more frequent classes. This issue is further pronounced in SSL methods, as they would use this biased model to obtain psuedo-labels (on the unlabeled data) during training. In this paper, we tackle this problem by attempting to select a balanced labeled dataset for SSL that would result in an unbiased model. Unfortunately, acquiring a balanced labeled dataset from a class imbalanced distribution in one shot is challenging. We propose BASIL (Balanced Active Semi-supervIsed Learning), a novel algorithm that optimizes the submodular mutual information (SMI) functions in a per-class fashion to gradually select a balanced dataset in an active learning loop. Importantly, our technique can be efficiently used to improve the performance of any SSL method. Our experiments on Path-MNIST and Organ-MNIST medical datasets for a wide array of SSL methods show the effectiveness of Basil. Furthermore, we observe that Basil outperforms the state-of-the-art diversity and uncertainty based active learning methods since the SMI functions select a more balanced dataset.

Raavi Gupta, Pranav Hari Panicker, Sumit Bhatia et al.

Large language models (LLMs), despite their remarkable text generation capabilities, often hallucinate and generate text that is factually incorrect and not grounded in real-world knowledge. This poses serious risks in domains like healthcare, finance, and customer support. A typical way to use LLMs is via the APIs provided by LLM vendors where there is no access to model weights or options to fine-tune the model. Existing methods to detect hallucinations in such settings where the model access is restricted or constrained by resources typically require making multiple LLM API calls, increasing latency and API cost. We introduce CONFACTCHECK, an efficient hallucination detection approach that does not leverage any external knowledge base and works on the simple intuition that responses to factual probes within the generated text should be consistent within a single LLM and across different LLMs. Rigorous empirical evaluation on multiple datasets that cover both the generation of factual texts and the open generation shows that CONFACTCHECK can detect hallucinated facts efficiently using fewer resources and achieves higher accuracy scores compared to existing baselines that operate under similar conditions. Our code is available here.

Piyush Singh Pasi, Karthikeya Battepati, Preethi Jyothi et al.

The problem of audio-to-text alignment has seen significant amount of research using complete supervision during training. However, this is typically not in the context of long audio recordings wherein the text being queried does not appear verbatim within the audio file. This work is a collaboration with a non-governmental organization called CARE India that collects long audio health surveys from young mothers residing in rural parts of Bihar, India. Given a question drawn from a questionnaire that is used to guide these surveys, we aim to locate where the question is asked within a long audio recording. This is of great value to African and Asian organizations that would otherwise have to painstakingly go through long and noisy audio recordings to locate questions (and answers) of interest. Our proposed framework, INDENT, uses a cross-attention-based model and prior information on the temporal ordering of sentences to learn speech embeddings that capture the semantics of the underlying spoken text. These learnt embeddings are used to retrieve the corresponding audio segment based on text queries at inference time. We empirically demonstrate the significant effectiveness (improvement in R-avg of about 3%) of our model over those obtained using text-based heuristics. We also show how noisy ASR, generated using state-of-the-art ASR models for Indian languages, yields better results when used in place of speech. INDENT, trained only on Hindi data is able to cater to all languages supported by the (semantically) shared text space. We illustrate this empirically on 11 Indic languages.

Karthika N J, Krishnakant Bhatt, Ganesh Ramakrishnan et al.

Translating technical terms into lexically similar, low-resource Indian languages remains a challenge due to limited parallel data and the complexity of linguistic structures. We propose a novel use-case of Sanskrit-based segments for linguistically informed translation of such terms, leveraging subword-level similarity and morphological alignment across related languages. Our approach uses character-level segmentation to identify meaningful subword units, facilitating more accurate and context-aware translation. To enable this, we utilize a Character-level Transformer model for Sanskrit Word Segmentation (CharSS), which addresses the complexities of sandhi and morpho-phonemic changes during segmentation. We observe consistent improvements in two experimental settings for technical term translation using Sanskrit-derived segments, averaging 8.46 and 6.79 chrF++ scores, respectively. Further, we conduct a post hoc human evaluation to verify the quality assessment of the translated technical terms using automated metrics. This work has important implications for the education field, especially in creating accessible, high-quality learning materials in Indian languages. By supporting the accurate and linguistically rooted translation of technical content, our approach facilitates inclusivity and aids in bridging the resource gap for learners in low-resource language communities.

Sravya Vardhani Shivapuja, Ashwin Gopinath, Ayush Gupta et al.

The data distribution in popular crowd counting datasets is typically heavy tailed and discontinuous. This skew affects all stages within the pipelines of deep crowd counting approaches. Specifically, the approaches exhibit unacceptably large standard deviation wrt statistical measures (MSE, MAE). To address such concerns in a holistic manner, we make two fundamental contributions. Firstly, we modify the training pipeline to accommodate the knowledge of dataset skew. To enable principled and balanced minibatch sampling, we propose a novel smoothed Bayesian binning approach. More specifically, we propose a novel cost function which can be readily incorporated into existing crowd counting deep networks to encourage bin-aware optimization. As the second contribution, we introduce additional performance measures which are more inclusive and throw light on various comparative performance aspects of the deep networks. We also show that our binning-based modifications retain their superiority wrt the newly proposed performance measures. Overall, our contributions enable a practically useful and detail-oriented characterization of performance for crowd counting approaches.

H S V N S Kowndinya Renduchintala, Sumit Bhatia, Ganesh Ramakrishnan

Instruction Tuning involves finetuning a language model on a collection of instruction-formatted datasets in order to enhance the generalizability of the model to unseen tasks. Studies have shown the importance of balancing different task proportions during finetuning, but finding the right balance remains challenging. Unfortunately, there's currently no systematic method beyond manual tuning or relying on practitioners' intuition. In this paper, we introduce SMART (Submodular data Mixture strAtegy for instRuction Tuning) - a novel data mixture strategy which makes use of a submodular function to assign importance scores to tasks which are then used to determine the mixture weights. Given a fine-tuning budget, SMART redistributes the budget among tasks and selects non-redundant samples from each task. Experimental results demonstrate that SMART significantly outperforms traditional methods such as examples proportional mixing and equal mixing. Furthermore, SMART facilitates the creation of data mixtures based on a few representative subsets of tasks alone and through task pruning analysis, we reveal that in a limited budget setting, allocating budget among a subset of representative tasks yields superior performance compared to distributing the budget among all tasks. The code for reproducing our results is open-sourced at https://github.com/kowndinya-renduchintala/SMART.

Suraj Racha, Prashant Harish Joshi, Utkarsh Maurya et al.

Large Language Models (LLMs) have shown remarkable capabilities for complex tasks, yet adaptation in medical domain, specifically mental health, poses specific challenges. Mental health is a rising concern globally with LLMs having large potential to help address the same. We highlight three primary challenges for LLMs in mental health - lack of high quality interpretable and knowledge grounded training data; training paradigms restricted to core capabilities, and evaluation of multi turn dialogue settings. Addressing it, we present oMind framework which includes training and aligning LLM agents for diverse capabilities including conversations; high quality ~164k multi-task SFT dataset, as a result of our generation pipeline based on Structured Knowledge retrieval, LLM based pruning, and review actions. We also introduce oMind-Chat - a novel multi turn benchmark dataset with expert annotated turn level and conversation level rubrics. Our diverse experiments on both core capabilities and conversations shows oMind LLMs consistently outperform baselines. oMind-LLM also shows significantly better reasoning with up to 80% win rate.

N J Karthika, Keerthana Suryanarayanan, Jahanvi Purohit et al.

We release Samasāmayik, a novel, meticulously curated, large-scale Hindi-Sanskrit corpus, comprising 92,196 parallel sentences. Unlike most data available in Sanskrit, which focuses on classical era text and poetry, this corpus aggregates data from diverse sources covering contemporary materials, including spoken tutorials, children's magazines, radio conversations, and instruction materials. We benchmark this new dataset by fine-tuning three complementary models - ByT5, NLLB and IndicTrans-v2, to demonstrate its utility. Our experiments demonstrate that models trained on the Samasamayik corpus achieve significant performance gains on in-domain test data, while achieving comparable performance on other widely used test sets, establishing a strong new performance baseline for contemporary Hindi-Sanskrit translation. Furthermore, a comparative analysis against existing corpora reveals minimal semantic and lexical overlap, confirming the novelty and non-redundancy of our dataset as a robust new resource for low-resource Indic language MT.

Manoj Balaji Jagadeeshan, Samarth Bhatia, Pretam Ray et al.

Recent advances in large language models (LLMs) have significantly improved natural language generation, including creative tasks like poetry composition. However, most progress remains concentrated in high-resource languages. This raises an important question: Can LLMs be adapted for structured poetic generation in a low-resource, morphologically rich language such as Sanskrit? In this work, we introduce a dataset designed for translating English prose into structured Sanskrit verse, with strict adherence to classical metrical patterns, particularly the Anushtub meter. We evaluate a range of generative models-both open-source and proprietary-under multiple settings. Specifically, we explore constrained decoding strategies and instruction-based fine-tuning tailored to metrical and semantic fidelity. Our decoding approach achieves over 99% accuracy in producing syntactically valid poetic forms, substantially outperforming general-purpose models in meter conformity. Meanwhile, instruction-tuned variants show improved alignment with source meaning and poetic style, as supported by human assessments, albeit with marginal trade-offs in metrical precision.

Dhruv Kudale, Badri Vishal Kasuba, Venkatapathy Subramanian et al.

Several recent deep learning (DL) based techniques perform considerably well on image-based multilingual text detection. However, their performance relies heavily on the availability and quality of training data. There are numerous types of page-level document images consisting of information in several modalities, languages, fonts, and layouts. This makes text detection a challenging problem in the field of computer vision (CV), especially for low-resource or handwritten languages. Furthermore, there is a scarcity of word-level labeled data for text detection, especially for multilingual settings and Indian scripts that incorporate both printed and handwritten text. Conventionally, Indian script text detection requires training a DL model on plenty of labeled data, but to the best of our knowledge, no relevant datasets are available. Manual annotation of such data requires a lot of time, effort, and expertise. In order to solve this problem, we propose TEXTRON, a Data Programming-based approach, where users can plug various text detection methods into a weak supervision-based learning framework. One can view this approach to multilingual text detection as an ensemble of different CV-based techniques and DL approaches. TEXTRON can leverage the predictions of DL models pre-trained on a significant amount of language data in conjunction with CV-based methods to improve text detection in other languages. We demonstrate that TEXTRON can improve the detection performance for documents written in Indian languages, despite the absence of corresponding labeled data. Further, through extensive experimentation, we show improvement brought about by our approach over the current State-of-the-art (SOTA) models, especially for handwritten Devanagari text. Code and dataset has been made available at https://github.com/IITB-LEAP-OCR/TEXTRON

Dhruv Kudale, Badri Vishal Kasuba, Venkatapathy Subramanian et al.

Table Structure Recognition (TSR) is vital for various downstream tasks like information retrieval, table reconstruction, and document understanding. While most state-of-the-art (SOTA) research predominantly focuses on TSR in English documents, the need for similar capabilities in other languages is evident, considering the global diversity of data. Moreover, creating substantial labeled data in non-English languages and training these SOTA models from scratch is costly and time-consuming. We propose TSR as a language-agnostic cell arrangement prediction and introduce SPRINT, Script-agnostic Structure Recognition in Tables. SPRINT uses recently introduced Optimized Table Structure Language (OTSL) sequences to predict table structures. We show that when coupled with a pre-trained table grid estimator, SPRINT can improve the overall tree edit distance-based similarity structure scores of tables even for non-English documents. We experimentally evaluate our performance across benchmark TSR datasets including PubTabNet, FinTabNet, and PubTables-1M. Our findings reveal that SPRINT not only matches SOTA models in performance on standard datasets but also demonstrates lower latency. Additionally, SPRINT excels in accurately identifying table structures in non-English documents, surpassing current leading models by showing an absolute average increase of 11.12%. We also present an algorithm for converting valid OTSL predictions into a widely used HTML-based table representation. To encourage further research, we release our code and Multilingual Scanned and Scene Table Structure Recognition Dataset, MUSTARD labeled with OTSL sequences for 1428 tables in thirteen languages encompassing several scripts at https://github.com/IITB-LEAP-OCR/SPRINT

Ayush Maheshwari, Atul Kumar Singh, Karthika NJ et al.

Lexicon or dictionary generation across domains has the potential for societal impact, as it can potentially enhance information accessibility for a diverse user base while preserving language identity. Prior work in the field primarily focuses on bilingual lexical induction, which deals with word alignments using mapping or corpora-based approaches. However, these approaches do not cater to domain-specific lexicon generation that consists of domain-specific terminology. This task becomes particularly important in specialized medical, engineering, and other technical domains, owing to the highly infrequent usage of the terms and scarcity of data involving domain-specific terms especially for low/mid-resource languages. In this paper, we propose a new model to generate dictionary words for $6$ Indian languages in the multi-domain setting. Our model consists of domain-specific and domain-generic layers that encode information, and these layers are invoked via a learnable routing technique. We also release a new benchmark dataset consisting of >75K translation pairs across 6 Indian languages spanning 8 diverse domains.We conduct both zero-shot and few-shot experiments across multiple domains to show the efficacy of our proposed model in generalizing to unseen domains and unseen languages. Additionally, we also perform a post-hoc human evaluation on unseen languages. The source code and dataset is present at https://github.com/Atulkmrsingh/lexgen.

Sarvesh Shashidhar, Ritik, Nachiketa Patil et al.

Direct Preference Optimisation (DPO) has emerged as a powerful method for aligning Large Language Models (LLMs) with human preferences, offering a stable and efficient alternative to approaches that use Reinforcement learning via Human Feedback. In this work, we investigate the performance of DPO using open-source preference datasets. One of the major drawbacks of DPO is that it doesn't induce granular scoring and treats all the segments of the responses with equal propensity. However, this is not practically true for human preferences since even "good" responses have segments that may not be preferred by the annotator. To resolve this, a 2-dimensional scoring for DPO alignment called 2D-DPO was proposed. We explore the 2D-DPO alignment paradigm and the advantages it provides over the standard DPO by comparing their win rates. It is observed that these methods, even though effective, are not robust to label/score noise. To counter this, we propose an approach of incorporating segment-level score noise robustness to the 2D-DPO algorithm. Along with theoretical backing, we also provide empirical verification in favour of the algorithm and introduce other noise models that can be present.

Anas Anwarul Haq Khan, Utkarsh Verma, Ganesh Ramakrishnan

We introduce DEEVISum (Distilled Early Exit Vision language model for Summarization), a lightweight, efficient, and scalable vision language model designed for segment wise video summarization. Leveraging multi modal prompts that combine textual and audio derived signals, DEEVISum incorporates Multi Stage Knowledge Distillation (MSKD) and Early Exit (EE) to strike a balance between performance and efficiency. MSKD offers a 1.33% absolute F1 improvement over baseline distillation (0.5%), while EE reduces inference time by approximately 21% with a 1.3 point drop in F1. Evaluated on the TVSum dataset, our best model PaLI Gemma2 3B + MSKD achieves an F1 score of 61.1, competing the performance of significantly larger models, all while maintaining a lower computational footprint. We publicly release our code and processed dataset to support further research.

Badri Vishal Kasuba, Dhruv Kudale, Venkatapathy Subramanian et al.

In recent years, the field of Handwritten Text Recognition (HTR) has seen the emergence of various new models, each claiming to perform competitively better than the other in specific scenarios. However, making a fair comparison of these models is challenging due to inconsistent choices and diversity in test sets. Furthermore, recent advancements in HTR often fail to account for the diverse languages, especially Indic languages, likely due to the scarcity of relevant labeled datasets. Moreover, much of the previous work has focused primarily on character-level or word-level recognition, overlooking the crucial stage of Handwritten Text Detection (HTD) necessary for building a page-level end-to-end handwritten OCR pipeline. Through our paper, we address these gaps by making three pivotal contributions. Firstly, we present an end-to-end framework for Page-Level hAndwriTTen TExt Recognition (PLATTER) by treating it as a two-stage problem involving word-level HTD followed by HTR. This approach enables us to identify, assess, and address challenges in each stage independently. Secondly, we demonstrate the usage of PLATTER to measure the performance of our language-agnostic HTD model and present a consistent comparison of six trained HTR models on ten diverse Indic languages thereby encouraging consistent comparisons. Finally, we also release a Corpus of Handwritten Indic Scripts (CHIPS), a meticulously curated, page-level Indic handwritten OCR dataset labeled for both detection and recognition purposes. Additionally, we release our code and trained models, to encourage further contributions in this direction.

Ayush Maheshwari, Ashim Gupta, Amrith Krishna et al.

We release Sāmayik, a dataset of around 53,000 parallel English-Sanskrit sentences, written in contemporary prose. Sanskrit is a classical language still in sustenance and has a rich documented heritage. However, due to the limited availability of digitized content, it still remains a low-resource language. Existing Sanskrit corpora, whether monolingual or bilingual, have predominantly focused on poetry and offer limited coverage of contemporary written materials. Sāmayik is curated from a diverse range of domains, including language instruction material, textual teaching pedagogy, and online tutorials, among others. It stands out as a unique resource that specifically caters to the contemporary usage of Sanskrit, with a primary emphasis on prose writing. Translation models trained on our dataset demonstrate statistically significant improvements when translating out-of-domain contemporary corpora, outperforming models trained on older classical-era poetry datasets. Finally, we also release benchmark models by adapting four multilingual pre-trained models, three of them have not been previously exposed to Sanskrit for translating between English and Sanskrit while one of them is multi-lingual pre-trained translation model including English and Sanskrit. The dataset and source code is present at https://github.com/ayushbits/saamayik.

H S V N S Kowndinya Renduchintala, Krishnateja Killamsetty, Sumit Bhatia et al.

A salient characteristic of pre-trained language models (PTLMs) is a remarkable improvement in their generalization capability and emergence of new capabilities with increasing model capacity and pre-training dataset size. Consequently, we are witnessing the development of enormous models pushing the state-of-the-art. It is, however, imperative to realize that this inevitably leads to prohibitively long training times, extortionate computing costs, and a detrimental environmental impact. Significant efforts are underway to make PTLM training more efficient through innovations in model architectures, training pipelines, and loss function design, with scant attention being paid to optimizing the utility of training data. The key question that we ask is whether it is possible to train PTLMs by employing only highly informative subsets of the training data while maintaining downstream performance? Building upon the recent progress in informative data subset selection, we show how we can employ submodular optimization to select highly representative subsets of the training corpora and demonstrate that the proposed framework can be applied to efficiently train multiple PTLMs (BERT, BioBERT, GPT-2) using only a fraction of data. Further, we perform a rigorous empirical evaluation to show that the resulting models achieve up to $\sim99\%$ of the performance of the fully-trained models. We made our framework publicly available at https://github.com/Efficient-AI/ingenious.

Duncan McElfresh, Sujay Khandagale, Jonathan Valverde et al.

Tabular data is one of the most commonly used types of data in machine learning. Despite recent advances in neural nets (NNs) for tabular data, there is still an active discussion on whether or not NNs generally outperform gradient-boosted decision trees (GBDTs) on tabular data, with several recent works arguing either that GBDTs consistently outperform NNs on tabular data, or vice versa. In this work, we take a step back and question the importance of this debate. To this end, we conduct the largest tabular data analysis to date, comparing 19 algorithms across 176 datasets, and we find that the 'NN vs. GBDT' debate is overemphasized: for a surprisingly high number of datasets, either the performance difference between GBDTs and NNs is negligible, or light hyperparameter tuning on a GBDT is more important than choosing between NNs and GBDTs. A remarkable exception is the recently-proposed prior-data fitted network, TabPFN: although it is effectively limited to training sets of size 3000, we find that it outperforms all other algorithms on average, even when randomly sampling 3000 training datapoints. Next, we analyze dozens of metafeatures to determine what properties of a dataset make NNs or GBDTs better-suited to perform well. For example, we find that GBDTs are much better than NNs at handling skewed or heavy-tailed feature distributions and other forms of dataset irregularities. Our insights act as a guide for practitioners to determine which techniques may work best on their dataset. Finally, with the goal of accelerating tabular data research, we release the TabZilla Benchmark Suite: a collection of the 36 'hardest' of the datasets we study. Our benchmark suite, codebase, and all raw results are available at https://github.com/naszilla/tabzilla.

Vishal Kaushal, Ganesh Ramakrishnan, Rishabh Iyer

Submodular functions are a special class of set functions which naturally model the notion of representativeness, diversity, coverage etc. and have been shown to be computationally very efficient. A lot of past work has applied submodular optimization to find optimal subsets in various contexts. Some examples include data summarization for efficient human consumption, finding effective smaller subsets of training data to reduce the model development time (training, hyper parameter tuning), finding effective subsets of unlabeled data to reduce the labeling costs, etc. A recent work has also leveraged submodular functions to propose submodular information measures which have been found to be very useful in solving the problems of guided subset selection and guided summarization. In this work, we present Submodlib which is an open-source, easy-to-use, efficient and scalable Python library for submodular optimization with a C++ optimization engine. Submodlib finds its application in summarization, data subset selection, hyper parameter tuning, efficient training and more. Through a rich API, it offers a great deal of flexibility in the way it can be used. Source of Submodlib is available at https://github.com/decile-team/submodlib.

Guttu Sai Abhishek, Harshad Ingole, Parth Laturia et al.

We present SPEAR, an open-source python library for data programming with semi supervision. The package implements several recent data programming approaches including facility to programmatically label and build training data. SPEAR facilitates weak supervision in the form of heuristics (or rules) and association of noisy labels to the training dataset. These noisy labels are aggregated to assign labels to the unlabeled data for downstream tasks. We have implemented several label aggregation approaches that aggregate the noisy labels and then train using the noisily labeled set in a cascaded manner. Our implementation also includes other approaches that jointly aggregate and train the model for text classification tasks. Thus, in our python package, we integrate several cascade and joint data-programming approaches while also providing the facility of data programming by letting the user define labeling functions or rules. The code and tutorial notebooks are available at https://github.com/decile-team/spear. Further, extensive documentation can be found at https://spear-decile.readthedocs.io/. Video tutorials demonstrating the usage of our package are available here. We also present some real-world use cases of SPEAR.

Nathan Beck, Durga Sivasubramanian, Apurva Dani et al.

With the goal of making deep learning more label-efficient, a growing number of papers have been studying active learning (AL) for deep models. However, there are a number of issues in the prevalent experimental settings, mainly stemming from a lack of unified implementation and benchmarking. Issues in the current literature include sometimes contradictory observations on the performance of different AL algorithms, unintended exclusion of important generalization approaches such as data augmentation and SGD for optimization, a lack of study of evaluation facets like the labeling efficiency of AL, and little or no clarity on the scenarios in which AL outperforms random sampling (RS). In this work, we present a unified re-implementation of state-of-the-art AL algorithms in the context of image classification via our new open-source AL toolkit DISTIL, and we carefully study these issues as facets of effective evaluation. On the positive side, we show that AL techniques are $2\times$ to $4\times$ more label-efficient compared to RS with the use of data augmentation. Surprisingly, when data augmentation is included, there is no longer a consistent gain in using BADGE, a state-of-the-art approach, over simple uncertainty sampling. We then do a careful analysis of how existing approaches perform with varying amounts of redundancy and number of examples per class. Finally, we provide several insights for AL practitioners to consider in future work, such as the effect of the AL batch size, the effect of initialization, the importance of retraining the model at every round, and other insights.

Atul Sahay, Anshul Nasery, Ayush Maheshwari et al.

Recently, unsupervised parsing of syntactic trees has gained considerable attention. A prototypical approach to such unsupervised parsing employs reinforcement learning and auto-encoders. However, no mechanism ensures that the learnt model leverages the well-understood language grammar. We propose an approach that utilizes very generic linguistic knowledge of the language present in the form of syntactic rules, thus inducing better syntactic structures. We introduce a novel formulation that takes advantage of the syntactic grammar rules and is independent of the base system. We achieve new state-of-the-art results on two benchmarks datasets, MNLI and WSJ. The source code of the paper is available at https://github.com/anshuln/Diora_with_rules.

Atul Sahay, Ayush Maheshwari, Ritesh Kumar et al.

Recursive neural networks (RvNN) have been shown useful for learning sentence representations and helped achieve competitive performance on several natural language inference tasks. However, recent RvNN-based models fail to learn simple grammar and meaningful semantics in their intermediate tree representation. In this work, we propose an attention mechanism over Tree-LSTMs to learn more meaningful and explainable parse tree structures. We also demonstrate the superior performance of our proposed model on natural language inference, semantic relatedness, and sentiment analysis tasks and compare them with other state-of-the-art RvNN based methods. Further, we present a detailed qualitative and quantitative analysis of the learned parse trees and show that the discovered linguistic structures are more explainable, semantically meaningful, and grammatically correct than recent approaches. The source code of the paper is available at https://github.com/atul04/Explainable-Latent-Structures-Using-Attention.

Krishnateja Killamsetty, Durga Sivasubramanian, Ganesh Ramakrishnan et al.

The great success of modern machine learning models on large datasets is contingent on extensive computational resources with high financial and environmental costs. One way to address this is by extracting subsets that generalize on par with the full data. In this work, we propose a general framework, GRAD-MATCH, which finds subsets that closely match the gradient of the training or validation set. We find such subsets effectively using an orthogonal matching pursuit algorithm. We show rigorous theoretical and convergence guarantees of the proposed algorithm and, through our extensive experiments on real-world datasets, show the effectiveness of our proposed framework. We show that GRAD-MATCH significantly and consistently outperforms several recent data-selection algorithms and achieves the best accuracy-efficiency trade-off. GRAD-MATCH is available as a part of the CORDS toolkit: \url{https://github.com/decile-team/cords}.

Sai Praneeth Reddy Sunkesula, Rishabh Dabral, Ganesh Ramakrishnan

Analyzing the interactions between humans and objects from a video includes identification of the relationships between humans and the objects present in the video. It can be thought of as a specialized version of Visual Relationship Detection, wherein one of the objects must be a human. While traditional methods formulate the problem as inference on a sequence of video segments, we present a hierarchical approach, LIGHTEN, to learn visual features to effectively capture spatio-temporal cues at multiple granularities in a video. Unlike current approaches, LIGHTEN avoids using ground truth data like depth maps or 3D human pose, thus increasing generalization across non-RGBD datasets as well. Furthermore, we achieve the same using only the visual features, instead of the commonly used hand-crafted spatial features. We achieve state-of-the-art results in human-object interaction detection (88.9% and 92.6%) and anticipation tasks of CAD-120 and competitive results on image based HOI detection in V-COCO dataset, setting a new benchmark for visual features based approaches. Code for LIGHTEN is available at https://github.com/praneeth11009/LIGHTEN-Learning-Interactions-with-Graphs-and-Hierarchical-TEmporal-Networks-for-HOI

Ayush Maheshwari, Oishik Chatterjee, KrishnaTeja Killamsetty et al.

The paradigm of data programming, which uses weak supervision in the form of rules/labelling functions, and semi-supervised learning, which augments small amounts of labelled data with a large unlabelled dataset, have shown great promise in several text classification scenarios. In this work, we argue that by not using any labelled data, data programming based approaches can yield sub-optimal performances, particularly when the labelling functions are noisy. The first contribution of this work is an introduction of a framework, \model which is a semi-supervised data programming paradigm that learns a \emph{joint model} that effectively uses the rules/labelling functions along with semi-supervised loss functions on the feature space. Next, we also study \modelss which additionally does subset selection on top of the joint semi-supervised data programming objective and \emph{selects} a set of examples that can be used as the labelled set by \model. The goal of \modelss is to ensure that the labelled data can \emph{complement} the labelling functions, thereby benefiting from both data-programming as well as appropriately selected data for human labelling. We demonstrate that by effectively combining semi-supervision, data-programming, and subset selection paradigms, we significantly outperform the current state-of-the-art on seven publicly available datasets. \footnote{The source code is available at \url{https://github.com/ayushbits/Semi-Supervised-LFs-Subset-Selection}}

Saeel Sandeep Nachane, Ojas Gramopadhye, Prateek Chanda et al. · ibm-research

In this paper, we propose a modified version of the MedQA-USMLE dataset, named MEDQA-OPEN, which contains open-ended medical questions without options to mimic clinical scenarios, along with clinician-approved reasoned answers. Additionally, we implement a prompt driven by Chain of Thought (CoT) reasoning, CLINICR, to mirror the prospective process of incremental reasoning, reaching a correct response to medical questions. We empirically demonstrate how CLINICR outperforms the state-of-the-art 5-shot CoT-based prompt (Liévin et al., 2022). We also present an approach that mirrors real-life clinical practice by first exploring multiple differential diagnoses through MCQ-CLINICR and subsequently narrowing down to a final diagnosis using MCQ-ELIMINATIVE. Finally, emphasizing the importance of response verification in medical settings, we utilize a reward model mechanism, replacing the elimination process performed by MCQ-ELIMINATIVE.

Prateek Chanda, Shrey Modi, Ganesh Ramakrishnan

In a distributed machine learning setting like Federated Learning where there are multiple clients involved which update their individual weights to a single central server, often training on the entire individual client's dataset for each client becomes cumbersome. To address this issue we propose $\methodprop$: a personalized coreset weighted federated learning setup where the training updates for each individual clients are forwarded to the central server based on only individual client coreset based representative data points instead of the entire client data. Through theoretical analysis we present how the average generalization error is minimax optimal up to logarithm bounds (upper bounded by $\mathcal{O}(n_k^{-\frac{2 β}{2 β+\boldsymbolΛ}} \log ^{2 δ^{\prime}}(n_k))$) and lower bounds of $\mathcal{O}(n_k^{-\frac{2 β}{2 β+\boldsymbolΛ}})$, and how the overall generalization error on the data likelihood differs from a vanilla Federated Learning setup as a closed form function ${\boldsymbol{\Im}}(\boldsymbol{w}, n_k)$ of the coreset weights $\boldsymbol{w}$ and coreset sample size $n_k$. Our experiments on different benchmark datasets based on a variety of recent personalized federated learning architectures show significant gains as compared to random sampling on the training data followed by federated learning, thereby indicating how intelligently selecting such training samples can help in performance. Additionally, through experiments on medical datasets our proposed method showcases some gains as compared to other submodular optimization based approaches used for subset selection on client's data.

Himanshu Patil, Geo Jolly, Ramana Raja Buddala et al.

Video Question Answering (VideoQA) models enhance understanding and interaction with audiovisual content, making it more accessible, searchable, and useful for a wide range of fields such as education, surveillance, entertainment, and content creation. Due to heavy compute requirements, most large visual language models (VLMs) for VideoQA rely on a fixed number of frames by uniformly sampling the video. However, this process does not pick important frames or capture the context of the video. We present a novel query-based selection of frames relevant to the questions based on the submodular mutual Information (SMI) functions. By replacing uniform frame sampling with query-based selection, our method ensures that the chosen frames provide complementary and essential visual information for accurate VideoQA. We evaluate our approach on the MVBench dataset, which spans a diverse set of multi-action video tasks. VideoQA accuracy on this dataset was assessed using two VLMs, namely Video-LLaVA and LLaVA-NeXT, both of which originally employed uniform frame sampling. Experiments were conducted using both uniform and query-based sampling strategies. An accuracy improvement of up to \textbf{4\%} was observed when using query-based frame selection over uniform sampling. Qualitative analysis further highlights that query-based selection, using SMI functions, consistently picks frames better aligned with the question. We opine that such query-based frame selection can enhance accuracy in a wide range of tasks that rely on only a subset of video frames.

Madhav Gupta, Vishak Prasad C, Ganesh Ramakrishnan

Subset selection-based methods are widely used to explain deep vision models: they attribute predictions by highlighting the most influential image regions and support object-level explanations. While these methods perform well in in-distribution (ID) settings, their behavior under out-of-distribution (OOD) conditions remains poorly understood. Through extensive experiments across multiple ID-OOD sets, we find that reliability of the existing subset based methods degrades markedly, yielding redundant, unstable, and uncertainty-sensitive explanations. To address these shortcomings, we introduce a framework that combines submodular subset selection with layer-wise, gradient-based uncertainty estimation to improve robustness and fidelity without requiring additional training or auxiliary models. Our approach estimates uncertainty via adaptive weight perturbations and uses these estimates to guide submodular optimization, ensuring diverse and informative subset selection. Empirical evaluations show that, beyond mitigating the weaknesses of existing methods under OOD scenarios, our framework also yields improvements in ID settings. These findings highlight limitations of current subset-based approaches and demonstrate how uncertainty-driven optimization can enhance attribution and object-level interpretability, paving the way for more transparent and trustworthy AI in real-world vision applications.

Durga Sivasubramanian, Lokesh Nagalapatti, Rishabh Iyer et al.

Federated Learning (FL) is used to learn machine learning models with data that is partitioned across multiple clients, including resource-constrained edge devices. It is therefore important to devise solutions that are efficient in terms of compute, communication, and energy consumption, while ensuring compliance with the FL framework's privacy requirements. Conventional approaches to these problems select a weighted subset of the training dataset, known as coreset, and learn by fitting models on it. Such coreset selection approaches are also known to be robust to data noise. However, these approaches rely on the overall statistics of the training data and are not easily extendable to the FL setup. In this paper, we propose an algorithm called Gradient based Coreset for Robust and Efficient Federated Learning (GCFL) that selects a coreset at each client, only every $K$ communication rounds and derives updates only from it, assuming the availability of a small validation dataset at the server. We demonstrate that our coreset selection technique is highly effective in accounting for noise in clients' data. We conduct experiments using four real-world datasets and show that GCFL is (1) more compute and energy efficient than FL, (2) robust to various kinds of noise in both the feature space and labels, (3) preserves the privacy of the validation dataset, and (4) introduces a small communication overhead but achieves significant gains in performance, particularly in cases when the clients' data is noisy.

Navyansh Mahla, Kshitij Sharad Jadhav, Ganesh Ramakrishnan

Large Language Models (LLMs) have demonstrated remarkable capabilities across various domains, particularly in task generalization for both text and vision data. While fine-tuning these models can significantly enhance their performance on specific downstream tasks, it often requires high-quality data that cannot be shared due to privacy concerns. Federated Learning (FL) offers a promising solution for collaborative training without direct data sharing. However, many parameter-efficient fine-tuning strategies for LLMs in FL, particularly those based on Low-Rank Adaptation (LoRA), face limitations. In this paper, we critically analyze the convergence and performance guarantees of popular FL frameworks utilizing LoRA, highlighting its suboptimal nature due to constrained subspace learning of low-rank matrices. This limitation hinders effective fine-tuning of LLMs in federated settings. Through rigorous analytical and empirical evaluations, we demonstrate that direct weight averaging outperforms LoRA-based strategies, leading to superior performance for fine-tuned models. Our comprehensive comparison unmasks inefficiencies in LoRA approaches and underscores the advantages of direct weight aggregation. We extend our analysis to low-rank gradient-based optimizers, such as GaLore, used during local training steps. Our findings show that GaLore along with direct-weight aggregation is a more effective approach, outperforming federated LoRA methods like FlexLoRA and FFA-LoRA across both text and image modalities. While privacy remains paramount in FL discourse, our focus is on assessing performance outcomes of federated fine-tuned models and evaluating various FL frameworks from both theoretical and empirical perspectives. Our findings advocate reassessing the reliance on LoRA within FL contexts, paving the way for more efficient training methodologies.

Kundeshwar Pundalik, Piyush Sawarkar, Nihar Sahoo et al.

Large Language Models (LLMs) have emerged as powerful general-purpose reasoning systems, yet their development remains dominated by English-centric data, architectures, and optimization paradigms. This exclusionary design results in structural under-representation of linguistically diverse regions such as India, where over 20 official languages and 100+ dialects coexist alongside phenomena like code-switching and diglossia. We introduce PARAM-1, a 2.9B parameter decoder-only, text-only language model trained from scratch with an explicit architectural and linguistic focus on Indian diversity. PARAM-1 is trained on a bilingual dataset consisting of only Hindi and English, constructed with a strong focus on fact-rich, high-quality content. It is guided by three core principles: equitable representation of Indic languages through a 25% corpus allocation; tokenization fairness via a SentencePiece tokenizer adapted to Indian morphological structures; and culturally aligned evaluation benchmarks across IndicQA, code-mixed reasoning, and socio-linguistic robustness tasks. By embedding diversity at the pretraining level-rather than deferring it to post-hoc alignment-PARAM-1 offers a design-first blueprint for equitable foundation modeling. Our results demonstrate that it serves as both a competent general-purpose model and a robust baseline for India-centric applications.

Suraj Racha, Prashant Joshi, Anshika Raman et al.

Mental health remains a challenging problem all over the world, with issues like depression, anxiety becoming increasingly common. Large Language Models (LLMs) have seen a vast application in healthcare, specifically in answering medical questions. However, there is a lack of standard benchmarking datasets for question answering (QA) in mental health. Our work presents a novel multiple choice dataset, MHQA (Mental Health Question Answering), for benchmarking Language models (LMs). Previous mental health datasets have focused primarily on text classification into specific labels or disorders. MHQA, on the other hand, presents question-answering for mental health focused on four key domains: anxiety, depression, trauma, and obsessive/compulsive issues, with diverse question types, namely, factoid, diagnostic, prognostic, and preventive. We use PubMed abstracts as the primary source for QA. We develop a rigorous pipeline for LLM-based identification of information from abstracts based on various selection criteria and converting it into QA pairs. Further, valid QA pairs are extracted based on post-hoc validation criteria. Overall, our MHQA dataset consists of 2,475 expert-verified gold standard instances called MHQA-gold and ~56.1k pairs pseudo labeled using external medical references. We report F1 scores on different LLMs along with few-shot and supervised fine-tuning experiments, further discussing the insights for the scores.

Aamod Thakur, Ajay Nagpal, Atharva Savarkar et al.

While model architecture and training objectives are well-studied, tokenization, particularly in multilingual contexts, remains a relatively neglected aspect of Large Language Model (LLM) development. Existing tokenizers often exhibit high token-to-word ratios, inefficient use of context length, and slower inference. We present a systematic study that links vocabulary size, pre-tokenization rules, and training-corpus composition to both token-to-word efficiency and model quality. To ground our analysis in a linguistically diverse context, we conduct extensive experiments on Indic scripts, which present unique challenges due to their high script diversity and orthographic complexity. Drawing on the insights from these analyses, we propose a novel algorithm for data composition that balances multilingual data for tokenizer training. Our observations on pretokenization strategies significantly improve model performance, and our data composition algorithm reduces the average token-to-word ratio by approximately 6% with respect to the conventional data randomization approach. Our tokenizer achieves more than 40% improvement on average token-to-word ratio against stateof-the-art multilingual Indic models. This improvement yields measurable gains in both model performance and inference speed. This highlights tokenization alongside architecture and training objectives as a critical lever for building efficient, scalable multilingual LLMs

N J Karthika, Maharaj Brahma, Rohit Saluja et al.

Tokenization plays a pivotal role in multilingual NLP. However, existing tokenizers are often skewed towards high-resource languages, limiting their effectiveness for linguistically diverse and morphologically rich languages such as those in the Indian subcontinent. This paper presents a comprehensive intrinsic evaluation of tokenization strategies across 17 Indian languages. We quantify the trade-offs between bottom-up and top-down tokenizer algorithms (BPE and Unigram LM), effects of vocabulary sizes, and compare strategies of multilingual vocabulary construction such as joint and cluster-based training. We also show that extremely low-resource languages can benefit from tokenizers trained on related high-resource languages. Our study provides practical insights for building more fair, efficient, and linguistically informed tokenizers for multilingual NLP.