←
Harish Tayyar Madabushi1,805 posts
Harish Tayyar Madabushi avatar

Harish Tayyar Madabushi

@harish

Senior Lecturer (Associate Professor) in Artificial Intelligence. Work on NLP (#NLProc) and Large Language Models

◷ Joined December 2008▧ 133 Photos and videos
Harish Tayyar Madabushi avatar
Harish Tayyar Madabushi@harish·Nov 7

Completely agree!

I’ve seen many people confused about the current state of AI-generation detection tools. I believe that all of the following can be true: - AI generation detection is not fully solved and will never be. - Currently there are AI detection tools with a very high accuracy. [1/3]

BathNLP avatar
⟳ Harish Tayyar Madabushi retweeted
BathNLP@BathNLP·Oct 22

BathNLP will be presenting four full papers at #𝗘𝗠𝗡𝗟𝗣𝟮𝟬𝟮𝟲 🎉 Our work covers a wide range of topics, from multilingual translation and linguistic grounding to text complexity and standards compliance. See our full list of papers below!

Post attachment
Dan Jurafsky avatar
⟳ Harish Tayyar Madabushi retweeted
Dan Jurafsky@jurafsky·Aug 20

To welcome you back to school, it's time for a new release of Speech and Language Processing! Our August 2026 edition finally has a chapter 1! Check it out here: web.stanford.edu/~jurafsky/s…

Haritz Puerto avatar
⟳ Harish Tayyar Madabushi retweeted
Haritz Puerto@HaritzPuerto·Aug 7

📢 Call for reviewers! We're organizing PALM, a #neurips2026 workshop on memory for AI systems, in Paris 🇫🇷 and we'd love your help! Have ≥1 paper at a major ML conference? Pls sign up🙏 🔗 palm-neurips-2026.github.io 📝 Register: forms.gle/QeVHzqjr7CkN5o777

Harish Tayyar Madabushi avatar
Harish Tayyar Madabushi@harish·20 Nov 2025

Happy to share that I’ve been promoted to Senior Lecturer in AI at the University of Bath (equivalent to Associate Professor). Grateful to colleagues, mentors, collaborators and students for all the support along the way. @UniofBath @BathNLP

Post attachment
Haritz Puerto avatar
⟳ Harish Tayyar Madabushi retweeted
Haritz Puerto@HaritzPuerto·30 Jul 2025

I’ll be presenting today at 11:00 in hall x5 booth 209 #acl2025nlp come and let’s talk about how to train with CoTs!

At first I was not sure🤔, but on second thought, I knew what to do!!!💡😃 📢 Diverse Chains of Thought help LLMs refine their Reasoning!! @haritzpuerto will be presenting our work at #acl2025nlp 🇦🇹 on Wednesday 30th at 11:00 #nlproc A 🧵👇

Post attachment
Haritz Puerto avatar
⟳ Harish Tayyar Madabushi retweeted
Haritz Puerto@HaritzPuerto·25 Jul 2025

Excited to present Diverse Chains of Thought at #acl2025nlp Do you have a dataset with more than one CoT/question? Do you want to train with that? Come to our poster to see how to make the most out of your data! 🗓️ Wednesday 30th 🕦11:00 📍Level 1 1.86 underline.io/events/485/post…

At first I was not sure🤔, but on second thought, I knew what to do!!!💡😃 📢 Diverse Chains of Thought help LLMs refine their Reasoning!! @haritzpuerto will be presenting our work at #acl2025nlp 🇦🇹 on Wednesday 30th at 11:00 #nlproc A 🧵👇

Post attachment
Harish Tayyar Madabushi avatar
Harish Tayyar Madabushi@harish·25 Jul 2025

We provide open access to our code, models, data, and results: 📽️Underline: underline.io/events/485/post… 📄Paper: aclanthology.org/2025.acl-lo… 💻 Code: github.com/UKPLab/acl2025-di… 🤗 Models: huggingface.co/collections/h… 📂 Data: tudatalib.ulb.tu-darmstadt.d… 🌐 Website: ukplab.github.io/acl2025-div… (9/🧵)

Harish Tayyar Madabushi avatar
Harish Tayyar Madabushi@harish·25 Jul 2025

We also observed that when we generate 3 CoTs, if the first 2 CoTs are ❌ and the 3rd is ✅, the model picks the last one! 🎉 This shows that DCoT is not an ensemble of CoTs and instead is doing self-correction 🎊 8/🧵

Harish Tayyar Madabushi avatar
Harish Tayyar Madabushi@harish·25 Jul 2025

Why does it work? DCoT attempts to generate subsequent correct CoTs. Maybe the first CoT is wrong ❌ (and the model doesn’t know it), but by trying to generate a second better CoT, the model may correct the first one ✅🤩 7/🧵

Harish Tayyar Madabushi avatar
Harish Tayyar Madabushi@harish·25 Jul 2025

Generating a second CoT is enough to achieve gains. Note that DCoT@1 remains the same as the vanilla CoT, i.e., training on DCoT is a better way to train an LLM if you have more than one CoT per question. (Both methods were trained with the same CoTs) 6/🧵

Post attachment
Harish Tayyar Madabushi avatar
Harish Tayyar Madabushi@harish·25 Jul 2025

What did we find? Fine-tuning LLMs with DCoT datasets significantly improves performance across all model sizes from 1.3B to 70B parameters. 🎉 5/🧵

Post attachment
Harish Tayyar Madabushi avatar
Harish Tayyar Madabushi@harish·25 Jul 2025

We train CoT and DCoT models with the CoTs. The only difference is that DCoT forces the model to generate them sequentially in a single inference step. With this, we wondered whether LMs can refine their reasoning on the go. 4/🧵

Post attachment
Harish Tayyar Madabushi avatar
Harish Tayyar Madabushi@harish·25 Jul 2025

We created a specialized DCoT dataset, where every question has multiple correct chains of thought. These alternative reasoning paths are all tied to the same answer, encouraging the model to explore diverse solutions simultaneously. 🤔➡️💡 3/🧵

Harish Tayyar Madabushi avatar
Harish Tayyar Madabushi@harish·25 Jul 2025

Traditional CoT methods focus on a single chain of reasoning to arrive at a solution. DCoT, on the other hand, requires models to generate ➡️multiple reasoning paths before producing a final answer, 🔄all in a single inference step. 2/🧵

Post attachment