Honored to share our ConvAI contributions to #EMNLP2024! Congrats to Ishika Agarwal, Sagnik Mukherjee, Sumuk 🤗, and Abhinav Chinta - we'll see you in Miami 🌴☀️

Feel GPT's style is not quite what you like? Wish that it could understand your preferences implicitly? Our latest work from ConvAI@UIUC, Unsupervised Human Preference Learning (now accepted to the EMNLP 2024 main conference), offers a novel, decoupled solution to the alignment

Congratulations Sumuk 🤗 Abhinav Chinta Vaibhav Sahai dilek hakkani-tur looking forward to your presentation in Miami 🎉

3 days left until #EMNLP2024!🌴ConvAI will be in Miami (DM if you'd like to meet up!) presenting the following works:

happening today at riverfront hall from 10:30 - 12:00! do drop by - we’ve done something quite clever and think you’ll be super excited!

Happening today at 10:30AM in Riverfront Hall!! Please drop by and check out our novel take on preference learning! #EMNLP

With a bit of prompting gemini 2.0 flash experimental was finally able to generate this. Tried with every other model and nothing even came close. “A horse riding an astronaut.”

we're launching 🤗 yourbench today, an open source tool for custom benchmarking and synthetic data generation from ANY of your documents. it's a big step towards improving how model evaluations work early access link in replies! (1/8)

🚀Our ICML 2025 paper introduces "Premise-Augmented Reasoning Chains" - a structured approach to induce explicit dependencies in reasoning chains. By revealing the dependencies within chains, we significantly improve how LLM reasoning can be verified. 🧵[1/n]

🚨 Paper Alert: “RL Finetunes Small Subnetworks in Large Language Models” From DeepSeek V3 Base to DeepSeek R1 Zero, a whopping 86% of parameters were NOT updated during RL training 😮😮 And this isn’t a one-off. The pattern holds across RL algorithms and models. 🧵A Deep Dive

🚀 Headed to #ICML2025 in Vancouver (July 13-19) ! We will present our paper in the poster session at East Exhibition Hall on Tuesday (15th) at 4:30 PM PDT. Happy to chat regarding reasoning, post-training and anything LLMs in general !

🚨Presenting our ICML paper “Premise Augmented Reasoning Chains” at 4:30PM PT today in the East Exhibition Hall A-B (E-2410). Come check out our work on improving error detection in COT using DAGs! Link to paper: abhinavchinta.com/parc/ #icml25

100% agree with will. building my first 3090 cluster with Abhinav Chinta was such a great learning experiment in sourcing cheap components from shenzhen, dealing with riser retiming issues, hacking power supplies together etc. just buy the service “advice” is toxic.

We introduce SuperCoder, the first work to successfully apply LLMs as superoptimizers for assembly code 🚀. Our RL-trained model achieves 95% correctness ✅ and 1.46× speedup ⚡over gcc -O3. 📄arxiv.org/pdf/2505.11480 💻github.com/Anjiang-Wei/Su… #LLM #Compilers #Code #Optimization