Extended reasoning has substantially improved multimodal large language models on tasks that require compositional and relational inference. A new arXiv preprint, "Think, Compress, Answer: Reinforcement Learning for Multimodal Grounding via Reasoning Compression," takes that improvement as its starting point.
Yet the paper argues that multimodal grounding imposes a distinct requirement: reasoning must be compressed. In other words, models need to think before answering, but the reasoning itself has to be condensed to fit the grounding task.
To address this, the paper outlines a reinforcement learning approach that follows a think-compress-answer structure: first reason, then compress that reasoning, then produce the grounded answer. The title frames this as a way to reconcile extended reasoning with the specific demands of multimodal grounding.