Good to see discussions around prompt sensitivity & challenges in reporting optimized baselines! Our POSIX work feels even more relevant now. Reporting a POSIX score alongside performance numbers would bring much-needed perspective. CC: Omar Khattab Rulin Shao Kowndinya Renduchintala

🚨 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!

📝 On the Effect of Instruction Tuning Loss on Generalization 👥 Anwoy Chatterjee Anwoy Chatterjee, Kowndinya Renduchintala Kowndinya Renduchintala, Sumit Bhatia, Tanmoy Chakraborty Tanmoy Chakraborty 💾 Code: github.com/kowndinya-rend… arxiv.org/abs/2507.07817 #MachineLearning #NLP #LLMs #TACL

🚀 LCS2 Sneak Peek Series for #EMNLP2025 🚀 📝 Do You Know About My Nation? Investigating Multilingual Language Models’ Cultural Literacy Through Factual Knowledge 👥 @Eshaant2000, Anwoy Chatterjee, Michael Saxon, Alon Albalak, William Yang Wang, Tanmoy Chakraborty 📌 Paper: aclanthology.org/2025.emnlp-mai…

🆕 at #EMNLP: Will a LM be better at answering questions about Germany in German than in French? Is national knowledge linguistically contingent? Interestingly, only for some multilingual models is this true. Aya knows China best in Chinese, but LLaMA's best in English always.

I will be virtually presenting our #EMNLP2025 paper at 10:30 am tomorrow, during the Computational Social Science 2 Oral Session. 🗓️ Nov 6 ⏲️ 10:30 AM 📌 A109 TL;DR: We introduce the XNationQA - a parallel multilingual and multicultural benchmark to evaluate the cultural

"Of course you think SAEs are feature detectors, you're a first year interpretability researcher, just posted to a safety community, LessWrong probably. You'll talk about circuits until you read Neel Nanda

Just because an AI wrote it, doesn’t mean I don’t mean it. Just because I used AI, doesn’t mean I don’t care. It’s like saying, “If you really cared about this contract, you’d do the budget math by hand.” Me being efficient with my time doesn’t mean I respect yours less. All

Repeat after me: LLM detectors don't work. AI image detectors are bound to fail. Despite what some companies claim, AI and non-AI data will eventually be indistinguishable

Unpopular opinion -- I don't think we should accept papers in top AI conferences whose contribution is mainly experimental but don't provide the data and tools to reproduce the experiments. I have seen several papers with flawed results, but without code there's no way of telling

stop acting like the em-dash is proof of AI. it’s called grammar — people have been using it to great effect for hundreds of years. read a book for once

Most researchers collect data first, then wonder why nobody reads their papers. Start with the story. The data comes after.

It feels so disappointing when you meticulously read the author rebuttals for the papers you’re reviewing, responding to them and engaging with the authors only to get ghosted by ALL the reviewers on your own paper as an author ACLRollingReview ACL 2026

1/1 accepted at #ACL2026 Mains. See you in San Diego 🇺🇸 More details and preprint soon …

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 ACL 2026