Inference-time scaling is great, but it only works if your model actually explores. RL training — by default — kills that exploration. If you are curious to learn why and what we did about it in about our recent ICLR 2026 paper, check out this 🧵!

🚨 Excited about ML/NLP and looking for a research internship on controlled text generation? Come work with us on advanced constraint processing in large language models at NAVER LABS Europe! ✨ Learn more and apply here: europe.naverlabs.com/job/2-7/

Internship: Advanced Constraint Processing in LLMs

The ability to control the outputs of Large Language Models is a central topic in NLP and Machine Learning, with applications to safety, trustworthiness, reasoning, etc. Following […]

europe.naverlabs.com