🌟 𝐀 𝐍𝐞𝐰 T𝐞𝐱𝐭𝐛𝐨𝐨𝐤 -- 𝐈𝐧𝐭𝐫𝐨𝐝𝐮𝐜𝐭𝐢𝐨𝐧 𝐭𝐨 𝐋𝐚𝐫𝐠𝐞 𝐋𝐚𝐧𝐠𝐮𝐚𝐠𝐞 𝐌𝐨𝐝𝐞𝐥𝐬 🌟 I am excited to share the release of my new textbook, 𝘐𝘯𝘵𝘳𝘰𝘥𝘶𝘤𝘵𝘪𝘰𝘯 𝘵𝘰 𝘓𝘢𝘳𝘨𝘦 𝘓𝘢𝘯𝘨𝘶𝘢𝘨𝘦 𝘔𝘰𝘥𝘦𝘭𝘴 (#LLMs) -- Perhaps the first textbook on LLMs. Target Audience: 👉 Students/beginners, Looking for a structured starting point to learn LLMs 👉 Teachers, planning to offer a course on LLMs 👉 Industry professional, seeking to deepen their understanding of LLMs Explore the Book: 🔗 Book Website: tanmoychak.com/llmbook/ 📑 Table of Contents: tanmoychak.com/llmbook/toc.p… 🛒 Available on Amazon: amazon.in/dp/936386474X/ Enhance Your Learning Experience: 👉 Slides & Lecture Videos: Chapter-wise resources -- lcs2-iitd.github.io/ELL881-A… 👉 Exercises & Solutions: Practice with detailed chapter exercises (solutions available on request). 👉 Upcoming @nptel_official Course: Starting January 2025! Preview here: onlinecourses.nptel.ac.in/no… Book Endorsement: 📖 Foreword by Prof. Tim Baldwin @eltimster 👏 Endorsements from Prof. Iryna Gurevych @IGurevych and Prof. Pushpak Bhattacharyya #LLMs #Textbook @iitdelhi @WileyIndiaPL @lcs2lab

🎉 Time to celebrate the acceptance of 4 papers at #NeurIPS2026 — all in the Main Track! The papers span multilingual LLM interpretability, knowledge distillation, and LLM personalization. Huge congratulations to all the students and collaborators! 👏 Details of the papers are here: lnkd.in/p/g8SdzPWH @NeurIPSConf @lcs2lab @iitdelhi

Congratulations to Sudipto Ghosh (@ScientificGhosh) on being awarded the prestigious @IndiaDST #INSPIRE #PhDFellowship! 🎓🎉 We are proud of his achievement and look forward to his continued contributions and success in the years ahead! ✨

🎉 Congratulations to Ayan Sengupta and Anwoy Chaterjee @anwoy_, PhD scholars from our lab, on being recognised among India’s Top 100 AI/ML Researchers through the #AmazonAI100 initiative! Proud to see their research contributions recognised at the national level! 🚀 #IITDelhi

Watch the full conversation with Prof. Tanmoy Chakraborty here: piped.video/watch?v=XyI9YlFj…

India has engineers building the AI frontier. Why are we still behind it? Prof. Tanmoy Chakraborty (@Tanmoy_Chak), IIT Delhi, on private investment, the talent we already have, and why India should keep building. A few highlights from our conversation.

I appeared on @TV9Bharatvarsh to discuss how AI could potentially devise plans to harm humans and what this means for AI safety. The full episode is here 👇 @iitdelhi @lcs2lab

I appeared on @BBCHindi yesterday to discuss AI का ख़तरा कितना असली है? (How real is the threat from AI?) piped.video/qHhn9ccBU5g?si=oOfL… via @YouTube (my part appeared around 31.20+ mins) @BBCBreaking @BBC @iitdelhi @lcs2lab

Multiple Positions Available: RA / Senior RA / Postdoc at our lab, @lcs2lab, @iitdelhi We are looking for motivated researchers to join @lcs2lab at @iitdelhi. Multiple full-time, in-person positions are available in the following research areas: • LLM Efficiency • Responsible LLMs • AI for Mental Healthcare Interested candidates are requested to carefully review the details and submit their applications through the Google Form below: forms.gle/wBxz8HcacceixNo96 Applications submitted via email will not be considered. However, you may reach out to me by email if you have any questions regarding the positions.

A two-day Indo-German Bilateral Workshop on Green AI for Healthcare and Mental Wellness was held at #IITDelhi on September 3-4, 2026. The workshop focused on Green AI for Healthcare, addressing the environmental and societal impacts of Generative AI with particular emphasis on sustainable, efficient models for real-world healthcare deployment. The following were the key objectives of the workship: • To advance sustainable and energy efficient AI solutions for healthcare and mental wellness. • To facilitate Indo-German knowledge exchange on Green AI, efficient AI models, and responsible deployment of AI in healthcare. • To explore innovative AI applications in areas such as medical imaging, clinical decision support, mental health, and multilingual healthcare. • To identify technical, infrastructural, regulatory, and ethical challenges in deploying AI across diverse healthcare settings. • To foster academic, clinical, and industry collaboration between India and Germany. • To establish a roadmap for future joint research, innovation, and capacity-building initiatives. The workshop was sponsored by the Indo-German Science and Technology Centre (IGSTC), a flagship joint initiative established by the Department of Science and Technology (DST), Government of India, and the Federal Ministry of Education and Research (BMBF), Germany. The workshop was organized by Prof. Tanmoy Chakraborty, IIT Delhi and Prof. Iryna Gurevych, TU Darmstadt.

What a wonderful panel discussion at GAI4HM'26, IIT Delhi, on "Will AI Co-Therapists Ever Be Feasible for Deployment in Developing and Culturally Diverse Countries? A Comparison of India and Germany." When I designed this topic, I never imagined it would spark such a thought-provoking and engaging discussion! Full credit goes to the wonderful panel members -- Prof. Anurag Agarwal (@AshokaUniv), Smriti Joshi (@wysabuddy), Dr Mona Duggal (@ICMRDELHI -NIRDHS), Dr Ragul Ganesh (JIPMER, Puducherry), Prof. Stefan Hofmann (Philipps-Universität Marburg), Prof. Oliver Grimm (University Hospital Frankfurt), Mona Sharma (Manorathi Foundation), Prof. Iryna Gurevych (TU Darmstadt), moderated by Dr Koushik Sinha Dev (@aiims_newdelhi)

Organising (w/ @IGurevych) GAI4HM -- the Indo-German Bilateral Workshop on Green AI for Healthcare & Mental Wellness at @iitdelhi -- an entirely in-person event, funded by @INDOGSTC! 🇩🇪 13 German speakers 🇮🇳 15+ Indian speakers 🏛️ 20+ institutions 💬 2 panel discussions Bringing together computer scientists, engineers, psychiatrists, psychologists, public health researchers, regulatory experts, industry representatives, and healthcare professionals -- a truly interdisciplinary forum at the intersection of Green AI, healthcare & mental wellness. @lcs2lab

We changed only the calibration set of a structured pruner and a pruned Llama-3.1-8B went from refusing 96% of harmful prompts to 13%. Unstructured pruning? Unaffected. Calibration-free pruning? Nothing to attack. Read the full article at - nitter.net/ayans007/status/209402… Work done at LCS2, IIT Delhi (Parmanu project) @Tanmoy_Chak @lcs2lab

I am delighted to announce the acceptance of 7 papers to EMNLP 2026 (Main: 4 and Findings: 3). Congratulations to the entire lab for the hard work!! The preprints of most of the papers are already available. The remaining will be uploaded soon. @lcs2lab @YardiScAI @iitdelhi @emnlpmeeting #EMNLP2026

Happy to share that both of our papers have been accepted at #EMNLP2026! 🎉 Thanks to all the co-authors: Hiba Arnaout, Clarissa W Ong, Juliet Bockhorst, Kate Sheehan, Rachael Moldow, @Tanmoy_Chak , Iryna Gurevych

** 𝐀𝐝𝐚𝐩𝐭𝐢𝐯𝐞 𝐀𝐠𝐞𝐧𝐭 𝐂𝐨𝐨𝐫𝐝𝐢𝐧𝐚𝐭𝐢𝐨𝐧 ** Multi-agent systems are powerful, but they can drastically multiply inference costs. Many existing systems rely on fixed or densely activated agent pipelines without adapting computation to each query: Which agents actually need to be consulted? How deep should the reasoning go? And when is communication worth its compute cost? Presenting 𝐆𝐑𝐀𝐃𝐄 -- Gated Routing and Adaptive Depth for Efficient Reasoning 🔗 Preprint: arxiv.org/abs/2607.10836 GRADE optimises multi-agent reasoning by: 🧠 𝐋𝐞𝐚𝐫𝐧𝐞𝐝 𝐂𝐨𝐨𝐫𝐝𝐢𝐧𝐚𝐭𝐢𝐨𝐧: We built a hierarchical system governed by lightweight gates that jointly manage agent selection, routing depth, communication, and pruning dynamically per query. ⚙️ 𝐂𝐨𝐆𝐑𝐏𝐎 𝐓𝐫𝐚𝐢𝐧𝐢𝐧𝐠: We adapt GRPO for collaborative settings with a novel, critic-free RL recipe that assigns a shared advantage signal to all participating agents and gates during a rollout. 🔄 𝐇𝐨𝐭-𝐒𝐰𝐚𝐩𝐩𝐚𝐛𝐥𝐞 𝐄𝐱𝐩𝐞𝐫𝐭𝐬: GRADE features an Expert Registry with per-agent calibration maps. You can swap out expert models at inference time using just 64 anchor queries, without retraining the gates. 🏆 At ~17B average active parameters, GRADE outperforms all baselines on GSM8K, GPQA, and MMLUPro -- beating the strongest baseline on MMLUPro by 4.8 points while using ~39% fewer active parameters. w/ @ScientificGhosh Do check out many more exciting works on small models and agentic coordination being developed as part of our mega project -- 𝐏𝐚𝐫𝐚𝐦𝐚𝐧𝐮 parmanu.lcs2.in/ @lcs2lab @iitdelhi #LLMEfficiency #MultiagentRounting #AgenticAI

*** Sharing our significant development in Model Interpretability *** Sparse autoencoder features are often interpretable, yet unreliable for steering. Our work shows why: knowing what a feature represents does not tell us how intervening on it will affect the model’s output across contexts. We introduce FEGA to study the downstream geometry of these effects. We find that consistent one-dimensional effects are rare across SAE variants. Value-like features, which encode relatively fixed concepts, tend to produce structured but multidimensional effects. Pointer-like features, which operate on context-supplied values through functions such as copying or rule application, predominantly produce diffuse effects. The key implication of our study is that an interpretable SAE feature need not provide a reliable direction for controlling model behaviour.

SAE features are often found to be interpretable but not useful for steering. SAEs are inherently trained for local re-construction, and most existing works study the local geometry and local effect of these features. In our latest work, we study the downstream geometry of SAE features and try to understand why most often they aren’t useful as stable steering directions. We introduce an analysis framework, FEGA, to study this. In the process, we were also able to distinguish between two classes of features: value-like ones encoding concepts, and pointer-like features encoding functions operating on context supplied values. We observe that consistent one-dimensional effects are rare across SAE variants. Value-like features more often produce structured low-dimensional effects, but usually across several directions, while pointer-like features predominantly produce diffuse effects. This means that a feature can be interpretable and causally relevant without behaving like one reusable steering vector. Check out our paper to learn more. Our project page also provides an interactive way to explore the nature of different features. 🌐 Project page: ukplab.github.io/FEGA/ 📄 Paper: arxiv.org/abs/2607.24645 💻 Code: github.com/UKPLab/FEGA It was a great collaboration with @UKPLab at @TUDarmstadt . Kudos to my amazing co-authors: Phu Gia Hoang, @Tanmoy_Chak , @IGurevych , and Subhabrata Dutta.