First Things First: Teaching LLM-Based Agents to Prioritize Must-Haves before Nice-to-Haves
This work introduces First Things First Reinforcement Learning (FTFRL) to help multimodal large language models prioritize mandatory user requirements over optional ones. By training models with a multi-objective reward system that checks for requirement classification, accuracy, and format adherence, the authors show that this approach significantly improves agents' ability to handle complex reasoning tasks and generalize better across different logic and math benchmarks.