RT @frankniujc: Hey this is me!.Our paper: Llama See, Llama Do: A Mechanistic Perspective on Contextual Entrainment and Distraction in LLMs….

RT @HaritzPuerto: 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!.

Join millions who have switched to Grok.

RT @HaritzPuerto: Excited to present Diverse Chains of Thought at #ACL2025NLP .Do you have a dataset with more than one CoT/question? Do yo….

@HaritzPuerto @UKPLab @BathNLP @IGurevych We provide open access to our code, models, data, and results:. 📽️Underline: 📄Paper: 💻 Code: 🤗 Models: 📂 Data: 🌐 Website: (9/🧵).

@HaritzPuerto @UKPLab @BathNLP @IGurevych 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/🧵.

@HaritzPuerto @UKPLab @BathNLP @IGurevych 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/🧵.

@HaritzPuerto @UKPLab @BathNLP @IGurevych 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/🧵

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

@HaritzPuerto @UKPLab @BathNLP @IGurevych 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/🧵

@HaritzPuerto @UKPLab @BathNLP @IGurevych 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/🧵.

@HaritzPuerto @UKPLab @BathNLP @IGurevych 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/🧵

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 🧵👇

RT @feralvam: The trial data has just been released to registered participants. There’s still time for your team to join! #emnlp2025 #nlproc.

📢Job Opportunity.Research Associate for Reasoning in LLMs, University of Bath, UK (Deadline 05 August 2025). We are looking to hire a highly motivated researcher to work on analysing reasoning in LLMs. For more information, see: 

RT @nedjmaou: The Cardiff #NLProc Workshop starts on Monday!.If you've registered, you should have received a confirmation email (from me).….

RT @StevenSchockae2: I am looking for a postdoctoral research associate to work on (LLM-based and neurosymbolic) reasoning for story unders….

RT @tylerl404: Happy to announce our journal paper on tongue twisters, Train and Constrain (TwistList 2.0), has now been officially publish….

RT @josephimperial_: 🚨 New global collaboration & dataset paper!. UniversalCEFR: Enabling Open Multilingual Research on Language Proficienc….

Here with @TomSFHaines , @releasethegecko and @davmoltisanti

Starting in 15 minutes! Looking forward to this talk by @DrGeofreyHinton #NLProc