New fine-tuning of language models: Match meaning, not tokens

Language models are usually trained to predict the next word, but that does not always lead to the best overall answers. We introduce energy-based fine-tuning, a new method that trains models to produce better full responses, leading to stronger results without the need for complex reward models or verifiers. Project: https://energy-based-fine-tuning.github.io Paper: https://arxiv.org/abs/2603.12248 GitHub: https://github.com/sjelassi/ebft_openrlhf This session aired on May 14, 2026, at Microsoft Research Forum, Season 2 Episode 4. Register for the series to hear about new releases: https://www.microsoft.com/en-us/research/event/microsoft-research-forum/?OCID=msr_researchforum_YTDescription Explore all previous episodes: https://aka.ms/researchforumYTplaylist