Researchers propose a new framework to improve the performance of spoken language models (SLMs) by addressing structural differences between speech and text. Current SLMs can generate textual responses from speech but struggle with instruction-following behavior and generalization across tasks. The authors argue that this is due to weak alignment between speech and text representations, which are not well-suited for each other's formats. They demonstrate the effectiveness of their framework through experiments on multiple benchmarks, achieving competitive performance against strong baselines. --- Why it matters: This matters because spoken language models are a key area of research in AI, with applications in voice assistants, transcription services, and more. Improving their performance could lead to better user experiences and more accurate speech-to-text capabilities. Source: https://arxiv.org/abs/2608.22908