Partha Talukdar

Proud to announce that we have released Updesh (beta), a synthetic, culturally-grounded instruction-tuning dataset covering 13 Indic languages, comprising of more than 8 million data points. The dataset is now available through AIKosh, the dataset platform of the IndiaAI mission (https://lnkd.in/gB7PskwC). This is our first contribution to AIKosh and hopefully the first of many to come, and we are grateful to have the opportunity to contribute to India's ambitious AI mission. Over the past year, our research has focused on multilingual instruction-tuning to build models that are more performant, safe and reliable on local languages, cultures and contexts. Synthetic data holds promise in bridging the data gap, particularly for improving Small Language Models, which could prove critical in India’s AI journey for running high-quality AI locally on affordable hardware. Why Updesh? Most public instruction-tuning corpora remain English-centric. Updesh provides ~8 million high-quality prompt / response pairs for Assamese, Bengali, Gujarati, Hindi, Kannada, Malayalam, Marathi, Nepali, Odia, Punjabi, Tamil, Telugu and Urdu. Key design choices - Designed to be used for both reasoning and generative tasks - Dual construction pipeline: selective translation of reasoning tasks + synthetic generation seeded with region-specific knowledge, making best use of data authored by native speakers in each language along with curated high-quality English datasets - Multi-stage quality pipeline with filtering, LLM-judge checks and model fine-tuning - full methodology and details will follow in a forthcoming tech report. This dataset release is the result of dedicated efforts led by our team of Research Fellows at the Multilingual NLP group at Microsoft Research India. We would also like to thank the IndiaAI mission team for helping us with the onboarding process on AIKosh, and to Abhishek Singh, Kavita Bhatia and Kalika Bali for their guidance and leadership. #OpenData #IndicNLP #Multilingual #LLM #IndiaAI #AIKosh

256

5 Comments

With growing investment in agentic AI, we need shared frameworks to set expectations and guide responsible progress. This paper (https://lnkd.in/gca56zaE) does exactly that, much like how staged autonomy levels helped advance self-driving when the problem was too complex to solve all at once. There are still open questions about how crisp the boundaries are between levels and how certification might work in practice, but these are exactly the kinds of challenges worth tackling now. A valuable step forward at the right time.

138

📝 Announcing our AAAI 2025 papers on LLM Prompting, De-biasing, and Causality 🔹 "𝐇𝐢𝐞𝐫𝐚𝐫𝐜𝐡𝐢𝐜𝐚𝐥 𝐏𝐫𝐨𝐦𝐩𝐭𝐢𝐧𝐠 𝐓𝐚𝐱𝐨𝐧𝐨𝐦𝐲: 𝐀 𝐔𝐧𝐢𝐯𝐞𝐫𝐬𝐚𝐥 𝐄𝐯𝐚𝐥𝐮𝐚𝐭𝐢𝐨𝐧 𝐅𝐫𝐚𝐦𝐞𝐰𝐨𝐫𝐤 𝐟𝐨𝐫 𝐋𝐚𝐫𝐠𝐞 𝐋𝐚𝐧𝐠𝐮𝐚𝐠𝐞 𝐌𝐨𝐝𝐞𝐥𝐬 𝐀𝐥𝐢𝐠𝐧𝐞𝐝 𝐰𝐢𝐭𝐡 𝐇𝐮𝐦𝐚𝐧 𝐂𝐨𝐠𝐧𝐢𝐭𝐢𝐯𝐞 𝐏𝐫𝐢𝐧𝐜𝐢𝐩𝐥𝐞𝐬" 🔹 Published at the AAAI 2025 CogSci-AI Neuro-Symbolic Bridge 🔹 In collaboration with the Indian Institute of Technology, Kharagpur 🔹We introduce a structured framework (HPT) to assess the cognitive demands of tasks for large language models (LLMs) by categorizing different prompting strategies in a hierarchical manner, improving LLM evaluation and performance through optimized prompting selection ✍🏼 Authors: Devichand Budagam, Ashutosh Kumar, Mahsa Khoshnoodi, Sankalp KJ, Vinija Jain, Aman Chadha 📝 Paper: https://lnkd.in/gwkcrACR 🔗 Poster: https://lnkd.in/eHeR2yv9 --- 🔹 "𝐔𝐧𝐛𝐨𝐱𝐢𝐧𝐠 𝐎𝐜𝐜𝐮𝐩𝐚𝐭𝐢𝐨𝐧𝐚𝐥 𝐁𝐢𝐚𝐬: 𝐆𝐫𝐨𝐮𝐧𝐝𝐞𝐝 𝐃𝐞𝐛𝐢𝐚𝐬𝐢𝐧𝐠 𝐨𝐟 𝐋𝐋𝐌𝐬 𝐰𝐢𝐭𝐡 𝐔𝐒 𝐋𝐚𝐛𝐨𝐫 𝐃𝐚𝐭𝐚" 🔹 Published at the AAAI 2025 Symposium on AI Trustworthiness and Risk Assessment for Challenged Contexts (ATRACC) 🔹 In collaboration with University of Maryland Baltimore County 🔹 We examine biases in LLMs related to gender, ethnicity, and occupation, evaluates their outputs against U.S. Bureau of Labor Statistics data, and proposes a simple yet effective debiasing method using real-world labor data to mitigate these biases. ✍🏼 Authors: Atmika Gorti, Aman Chadha, Manas Gaur 📝 Paper: https://lnkd.in/eNSq8PXx --- 🔹 "𝐂𝐚𝐮𝐬𝐞 𝐚𝐧𝐝 𝐄𝐟𝐟𝐞𝐜𝐭: 𝐂𝐚𝐧 𝐋𝐚𝐫𝐠𝐞 𝐋𝐚𝐧𝐠𝐮𝐚𝐠𝐞 𝐌𝐨𝐝𝐞𝐥𝐬 𝐓𝐫𝐮𝐥𝐲 𝐔𝐧𝐝𝐞𝐫𝐬𝐭𝐚𝐧𝐝 𝐂𝐚𝐮𝐬𝐚𝐥𝐢𝐭𝐲?" 🔹 Published at the AAAI 2025 Symposium on AI Trustworthiness and Risk Assessment for Challenged Contexts (ATRACC) 🔹 In collaboration with Carnegie Mellon University, University of North Texas, Rensselaer Polytechnic Institute, and University of Massachusetts Amherst 🔹 We introduce the Context-Aware Reasoning Enhancement with Counterfactual Analysis (CARE-CA) framework to enhance causal reasoning and explainability in LLMs. It integrates explicit causal detection using ConceptNet and counterfactual statements, alongside implicit causal detection through LLMs, and introduces CausalNet, a new dataset for further research in causal reasoning. ✍🏼 Authors: Swagata Ashwani, Kshiteesh Hegde, Nishith Reddy Mannuru, Mayank Jindal, Dushyant Singh Sengar, Krishna Chaitanya Rao, Ph.D, Dishant Banga, Vinija Jain, Aman Chadha 📝 Paper: https://lnkd.in/ejziinfm Association for the Advancement of Artificial Intelligence (AAAI) #artificialintelligence #genai #research

117

1 Comment

💥 ICML-25 Paper Alert! Introducing the JETTS benchmark, Judge Evaluation for Test-Time Scaling ⏲️ LLM-judges specialize in evaluating responses generated by other LLMs. Test-time scaling requires multiple model responses to be evaluated for the best one. So, can LLM-judges help test-time scaling? Check this thread to find the answer and some interesting analysis👇 Repo: https://bit.ly/3YteE6Z X: https://lnkd.in/gXvQsd4C Paper: https://bit.ly/4jtRGVM

105

Excited to share our latest work on repository-level code summarization using local LLMs for business applications!! In large-scale software development, understanding the intent behind complex codebases, especially for business applications, can be a major hurdle. Existing summarization methods struggle with large code artifacts and fail to capture the business context. We introduce a two-step approach that first summarizes smaller code units using syntax analysis and then aggregates them to generate higher-level file/package summaries keeping business context into account. Check out the full paper (to appear at LLM4Code@ICSE 2025) at https://lnkd.in/gKwfGTpm ! 🔍💡 Co-authors: Nilesh Dhulshette Vinay Kulkarni #CodeSummarization #LLM #LLM4Code #BusinessApplications #Research #Telecommunications #OSS #BSS #IEEE #TCS #TCSResearch

55

4 Comments

Thrilled to announce that our paper, “Why Attention Fails: A Taxonomy of Faults in Attention-Based Neural Networks,” by Sigma Jahan, Saurabh Singh Rajput , Tushar Sharma, and Masud Rahman, PhD, has been accepted for presentation at the 48th IEEE/ACM International Conference on Software Engineering (ICSE 2026) — the world’s premier venue in software engineering research (CORE A* ranked). In this paper, we conduct the first systematic investigation of faults in attention-based neural networks, covering 555 real-world failure cases from 96 projects across platforms such as GitHub, Hugging Face, and Stack Overflow. We develop a seven-category taxonomy of attention-specific faults and identify key fault patterns such as masking errors, context misalignment, projection faults, and cache inconsistencies. Our taxonomy bridges software engineering and AI reliability, providing a foundation for future debugging and fault diagnosis in modern transformer-based architectures. Preprint: https://lnkd.in/e3K9vw2f #DalhousieUniversity #RAISELab #ICSE2026 #AI #SoftwareEngineering #Reliability #AttentionMechanisms #MachineLearning #NeuralNetworks

102

3 Comments

The AICON 2025 event at SRM University focused on AI, Generative AI, and drug design for newcomers. My talk focused on a foundational understanding of the data-to-information-to-knowledge pipeline, explaining how to build AI models and perform inference and fine-tuning. A real-world case study illustrated how to identify epilepsy targets by mining existing scientific literature based on recent research. The presentation also showcased the power of modern tools, such as Google Gemini and molecular informatics platforms, to accelerate research, demonstrating how six months' worth of work could potentially be condensed into six minutes, particularly for data generation of molecular structures from names and its applications in QSAR and molecular docking. The core message emphasized empowering researchers and faculty to learn and use AI tools effectively for their research without losing enthusiasm (Never Give Up!). Four rules of success (Hon. Dr. APJ Kalam) 1). Think Big, 2). Work Hard 3. Keep Learning and 4. Never Give up.A few slides and data sensitive to challenges were removed from the attached presentation (Shared for Educational purposes only). Pharmaceutical Research and Academic Collaboration Welcome. #research #chemoinformatics #bioinformatics #drugdesign #ai #LLM #epilepsy #seizures #education #programming #pharma

55

5 Comments

I recently had the privilege of reviewing the book "Machine Learning Assisted Evolutionary Multi- and Many-Objective Optimization" by Saxena et al. for the journal Genetic Programming and Evolvable Machines. The book explores the potential of combining machine learning and evolutionary algorithms to solve complex optimization problems. Overall, this book is a valuable resource for researchers and practitioners interested in the intersection of machine learning and evolutionary optimization. It provides a solid foundation for understanding the state-of-the-art in this emerging field and outlines promising research directions for the future. Here is the link for my review: https://lnkd.in/dMcVNJ33 #machinelearning #evolutionaryoptimization #bookreview

58

2 Comments

Tomorrow at #SSDBM2025: Tanya will be presenting our vision for "OmniMesh: Addressing Findability Challenges in Distributed Nature Data Repositories"! This is a joint effort with Wei-Lun (Harry) Chao, Hilmar Lapp, Carl Boettiger, and Rongjun Qin. Findability is a critical impediment to biodiversity work: datasets are scattered and hard to find across many siloed repositories. At the same time, biodiversity data is inherently multimodal, making annotation a bottleneck. OmniMesh combines embeddings and zero-shot models to stitch these silos together with a lightweight, standards-based search layer. #multimodal #search #FAIR #ssdbm

44

2 Comments

AI citation hallucinations have become a growing concern across research and publishing. When AI systems generate convincing but completely fabricated citations, it poses real risks to research integrity. So I decided to build a hobby project around it—and use it as an opportunity to explore some modern Python tools I'd been meaning to try. The result: a tool for Simon Willison's llm app that verifies academic citations against Crossref, catching fake DOIs before they get incorporated into GenAI responses. This was as much a learning exercise as anything else—getting hands-on with uv for package management, experimenting with GitHub Copilot as a pair programmer, and diving into plugin architecture patterns. The tool itself is simple but addresses a growing problem. With AI writing tools becoming more common in academic and professional settings, we need practical ways to maintain research integrity. Thanks to Gareth Wright for his helpful guidance on package publishing!

100

7 Comments