🌟 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

What makes a good teacher? On-policy distillation has spent the year reinventing loss functions to fix problems that come from one source: the teacher doesn't know the student. New article on why every popular OPD loss has an unbounded advantage and why the fix isn't another loss. Read the full article at - x.com/ayans007/status/206541… w/ @Tanmoy_Chak @lcs2lab #LLM #KnowledgeDistillation

Markovian ODE-guided scoring can assess the quality of offline reasoning traces in language models Arghodeep Nandi, Ojasva Saxena, Tanmoy Chakraborty arxiv.org/abs/2603.01580 [𝚌𝚜.𝙲𝙻]

A reporting checklist for large language models in behavioural science dlvr.it/TSxtMM

Check out our recent work published in Nature Human Behaviour on responsible LLM checklist: nature.com/articles/s41562-0… @NatureHumBehav

This is huge. Our PEFT method, MonteCLoRA, has been merged with @huggingface. Do use it. Believe me. It is much much better than LoRA in terms of efficiency and stability.

Time to celebrate acceptance of two papers in 𝐈𝐂𝐌𝐋'26, including one 𝐒𝐩𝐨𝐭𝐥𝐢𝐠𝐡𝐭 (top 2.2%) 🎉 👉 Polaris: Coupled Orbital Polar Embeddings for Hierarchical Concept Learning 📔 arxiv.org/pdf/2605.00265 ✨ Introduces Polaris -- a hyperspherical embedding framework that decouples semantics from hierarchy using orbital geometry, uncertainty-aware learning, and efficient retrieval. 👉 Linguistic Properties and Model Scale in Brain Encoding: From Small to Compressed Language Models (#𝐒𝐩𝐨𝐭𝐥𝐢𝐠𝐡𝐭) 📔 arxiv.org/pdf/2602.07547 ✨ Shows that compact ~3B models can match much larger LLMs in brain alignment, with robustness even under compression. Grateful to all collaborators and students for the amazing work! 🚀 @icmlconf @lcs2lab @iitdelhi #ICML26

🇧🇷 #LCS2 goes to #Rio 🇧🇷 Presenting our paper where we move beyond memoryless personalization → modeling user preferences as action-conditioned geometric walks with memory for better, user-aligned summaries. See you at #Riocentro 🚀 #Personalization #RepresentationLearning

🚨 CLEF 2026 - CheckThat! Lab We are excited to announce that we are organising a task at this year’s CheckThat! Lab, which extends the fact-checking pipeline with a new task focused on an important step in professional fact-checking: generating full fact-checking articles 📰

🚨 Submissions are now open for the Conference for AI Scientists (CAISc) 2026, co-organised by Lossfunk and @bitspilaniindia. Submit to probe what happens when AI systems drive scientific discovery. Submissions are open until May 15! Here is everything you need to know 🧵

An AI system MUST be the primary author: that's the only rule! Thrilled to be co-organizing this pioneering conference CAISc 2026! Send in your AI-driven research by May 15th.. @bitspilaniindia @ramgopal_rao @murari_ai @Tanmoy_Chak @palashiitkgp

Six papers from our lab have been accepted for publication in #ACL2026. The papers cover topics including Interpretability, empowering small VLMs with advanced tool calling, LLM personalisation, and different benchmarking. #nlproc @aclmeeting

I strongly condemn and protest against rejecting a paper from ACL with such a justification. If I am not mistaken, "Findings" started with the motivation of accommodating such borderline "good" papers. I don’t see any reason behind such a justification, given that ACL does not have any venue constraints (runs in hybrid mode). #ACL2026 #NLProc @aclmeeting

Our newly introduced 𝐆𝐔𝐈𝐃𝐄-𝐋𝐋𝐌 -- A reporting checklist for using LLMs in behavioral & social science. Massive collaborative effort led by @stfeuerriegel.

2/ The organising committee for CAISc 2026 is led by @paraschopra, @dhruvtrehan9, and @gargdhruv36. We are glad to have @Tanmoy_Chak (IIT Delhi), Palash Goyal (Google Research), Dr Mohan Kankanhalli (NUS AI Institute), Shirish Karande (TCS Research) on our steering committee, and @murari_ai and Pratik Narang as our Program Committee Chairs. Additionally, our program committee for final human review spans CS, Mathematics, electrical engineering, and not just ML.

Our new study on interpretability explains -- 𝐭𝐡𝐞 𝐏𝐡𝐲𝐬𝐢𝐜𝐬 𝐨𝐟 𝐊𝐕 𝐂𝐚𝐜𝐡𝐞 𝐂𝐨𝐦𝐩𝐫𝐞𝐬𝐬𝐢𝐨𝐧 𝐟𝐨𝐫 𝐋𝐋𝐌𝐬 Pre-print: arxiv.org/abs/2603.01426 As context lengths continue to grow, the KV cache has become the primary memory bottleneck during inference. While many compression techniques report impressive memory savings with minimal drops in benchmark accuracy, we asked a more structural question: 👉 𝘞𝘩𝘢𝘵 𝘢𝘤𝘵𝘶𝘢𝘭𝘭𝘺 𝘩𝘢𝘱𝘱𝘦𝘯𝘴 𝘵𝘰 𝘢𝘵𝘵𝘦𝘯𝘵𝘪𝘰𝘯 𝘢𝘯𝘥 𝘳𝘦𝘢𝘴𝘰𝘯𝘪𝘯𝘨 𝘸𝘩𝘦𝘯 𝘸𝘦 𝘤𝘰𝘮𝘱𝘳𝘦𝘴𝘴 𝘵𝘩𝘦 𝘒𝘝 𝘤𝘢𝘤𝘩𝘦? We frame KV compression as a 𝐜𝐨𝐧𝐭𝐫𝐨𝐥𝐥𝐞𝐝 𝐩𝐞𝐫𝐭𝐮𝐫𝐛𝐚𝐭𝐢𝐨𝐧 𝐨𝐟 𝐭𝐨𝐤𝐞𝐧-𝐥𝐞𝐯𝐞𝐥 𝐫𝐨𝐮𝐭𝐢𝐧𝐠 𝐢𝐧 𝐬𝐞𝐥𝐟-𝐚𝐭𝐭𝐞𝐧𝐭𝐢𝐨𝐧. Rather than evaluating only final task accuracy, we design synthetic datasets to probe: (1) Multi-entity tracking, (2) Coreference resolution, and (3) Multi-hop reasoning. This setup allows us to disentangle three critical dimensions: Information Retention, Accessibility, and Utilisation. Our findings reveal an interesting pattern: 👉 𝐌𝐨𝐝𝐞𝐫𝐚𝐭𝐞 𝐜𝐨𝐦𝐩𝐫𝐞𝐬𝐬𝐢𝐨𝐧 often preserves surface-level accuracy despite substantial internal representational degradation — suggesting significant redundancy in current models. 👉 𝐍𝐞𝐚𝐫 𝐞𝐱𝐭𝐫𝐞𝐦𝐞 𝐜𝐨𝐦𝐩𝐫𝐞𝐬𝐬𝐢𝐨𝐧, we observe a sharp "safety cliff" in hallucinations, driven by global erasure of answer-critical tokens. 👉 We also uncover a second failure mode -- 𝐫𝐞𝐩𝐫𝐞𝐬𝐞𝐧𝐭𝐚𝐭𝐢𝐨𝐧𝐚𝐥 𝐫𝐢𝐠𝐢𝐝𝐢𝐭𝐲 -- where tokens remain present, but routing flexibility collapses. These results suggest that evaluating compression solely through downstream accuracy can mask stronger structural effects on reasoning. Understanding these internal dynamics is crucial as we move toward longer-context and more memory-efficient LLMs. Brilliant work by Ayan Sengupta and Samhruth Ananthanarayanan. #ScienceofLLMs #Interpretability #KVCache #ModelCompression

📢 New Media Feature 📢 We are pleased to share that Prof. Tanmoy Chakraborty @Tanmoy_Chak was recently invited to Insight, a program on @sansad_tv, to speak on the growing challenges posed by #AIGenerated synthetic content. #ResponsibleAI #PublicPolicy youtu.be/sxaLdmlVRtY?si=Uf4W…

On the eve of the India AI Impact Summit, I have recently joined a podcast on AI's impact and the road ahead. youtube.com/watch?v=t32cHAYj…

🎉 New Paper Alert 🎉 We're excited to share that our paper "Here for You: A Co-Designed Mental Health Screening App for Indian University Students " has been accepted for publication in #JMIR #FormativeResearch! 🌟 @Tanmoy_Chak preprints.jmir.org/preprint/…

Happy to announce that our paper has been accepted to #ICLR2026! 🎉 📜 Beyond Markovian Drifts: Action-Biased Geometric Walks with Memory for Personalized Summarization 👥 Parthiv Chatterjee, Asish Batha, Tashvi Patel, @sourish_rygbee, @Tanmoy_Chak Congratulations to all authors!

Amidst the #ICLR26 chaos, at least one ray of hope finally materialised. No. I am not coming to Rio 🙃 @iclr_conf @sourish_rygbee @lcs2lab @iitdelhi