💡 By reframing #TaxonomyExpansion as lineage-oriented #Reasoning, #LORex moves beyond brittle pipelines. Structured knowledge deserves structured thinking. #KnowledgeGraphs #NLProc.

#LORex outperforms 12 strong baselines across 4 benchmarks using TEMPORA’s lineage-aware ranking.📈 +12% taxonomy expansion accuracy.📐 +5% Wu & Palmer similarity.🎯 +21.3% Hit@k .Without any LLM fine-tuning, #LORex is efficient, scalable, and plug-and-play across domains.

🧠 Hierarchies drive recommender systems, IR, and semantic search, but taxonomy expansion remains noisy and error-prone. #LORex ranks candidates via TEMPORA. It chunks candidates for LLMs to reason over parent-child links with contextual verification.

📝 Rank, Chunk and Expand: Lineage-Oriented Reasoning for Taxonomy Expansion.👥 Sahil Mishra @sahilmishra0012, Kumar Arjun, Tanmoy Chakraborty @Tanmoy_Chak .📌 Paper: 💾 Code:

🚨 New #ACL2025Findings Paper 🚨.How can we accurately grow taxonomies without brittle heuristics or finetuned LLMs? Our new work introduces LORex, a plug-and-play framework that combines discriminative ranking with generative reasoning to expand taxonomies scalably & faithfully.

RT @Tanmoy_Chak: I am attending #ACL2025. Happy to catch up to discuss our lab's work (@lcs2lab) and opportunities at @iitdelhi. I also lo….

📌 Can we trust a model that gets the right output for the wrong reason? As KD becomes standard for compressing LMs, understanding how knowledge transfers is vital. Our findings urge a rethink of trust and interpretability in distilled models. #ACL2025 #NLProc #TrustworthyAI.

🧠 Distilling into larger students yields limited gains. But more crucially, many students produce the correct answers without replicating the teacher’s reasoning. This exposes a key dissonance: KD boosts accuracy, but not necessarily reasoning fidelity.

📈 We conduct a large-scale study distilling 7B–14B LMs into 0.5B–7B models on 14 complex zero-shot reasoning tasks. KD boosts small model performance by up to 10%. However, teacher size ≠ student performance. Task-specific expertise matters more.

📝 On the Generalization vs Fidelity Paradox in Knowledge Distillation.👥 Suhas Kamasetty Ramesh, Ayan Sengupta @ayans007, Tanmoy Chakraborty @Tanmoy_Chak .📌 Paper: 💾 Code:

📢 New #ACL2025Findings Paper 📢.Can smaller language models learn how large ones reason, or just what they conclude? Our latest paper in #ACLFindings explores the overlooked tension in #KnowledgeDistillation - generalization vs reasoning fidelity. #NLProc

☑️ HiPPrO outperforms LLMs like GPT-3.5 and GPT-4 in producing more aligned, controllable, and contextually appropriate counterspeech, while remaining #lightweight and #efficient.

🧠 Using a two-stage architecture, HiPPrO generates responses that reflect both the speaker's intent (like questioning or denouncing) and the emotional tone (like anger or joy), making it far more human-like and impactful in tone-sensitive contexts like #hatemitigation.

📝 Counterspeech the Ultimate Shield! Multi-Conditioned Counterspeech Generation through Attributed Prefix Learning .👥 @aswiniandpunter, @anilHHH87, @Tanmoy_Chak .💾 Code: #ResponsibleAI #HateSpeech #ContentModeration.

🚀 #ACL2025 Sneak Peek 🚀.As online hate continues to challenge platforms and communities, our lab takes a meaningful step forward with HiPPrO, a new framework for generating controlled, constructive #counterspeech 🛡️.#NLProc #AIResearch

RT @lcs2lab: 🚨 New #TACL Paper Alert 🚨.We explore a crucial question in instruction tuning: should we weight prompt and response tokens dif….

📝 On the Effect of Instruction Tuning Loss on Generalization.👥 Anwoy Chatterjee @anwoy_, Kowndinya Renduchintala @KowndinyaR, Sumit Bhatia, Tanmoy Chakraborty @Tanmoy_Chak.💾 Code: #MachineLearning #NLP #LLMs #TACL.

🚨 New #TACL Paper Alert 🚨.We explore a crucial question in instruction tuning: should we weight prompt and response tokens differently in the loss function?.Introducing Weighted Instruction Tuning - a simple idea that boosts generalization by up to +6.55% across 5 benchmarks!

📜 Step-by-Step Unmasking for Parameter-Efficient Fine-tuning of Large Language Models.👥 @AradhyeAgarwal, Ayan Sengupta, Suhas K Ramesh, @Tanmoy_Chak .💾 Code: 📌 Paper: #LLMs #NLProc #PEFT #LoRA #Adapters #AIResearch.

🚨 New Paper Alert 🚨. Thrilled to share that our latest work has been accepted to #TACL! 🎊. We introduce ID3, a dynamic #PEFT method that unmasks parameters incrementally, guided by a novel heuristic based on magnitude and gradient information. 🧠. 🔗