Researchers have developed a new framework called FiLoRA that allows them to control how multimodal foundation models rely on different internal feature pathways. This is achieved through instruction-conditioned parameter adaptation, which enables the model to selectively amplify or suppress specific features based on natural language instructions. The authors evaluated FiLoRA in various settings and found that it induces consistent and interpretable shifts in feature reliance without altering task semantics. --- Why it matters: This matters because it provides a new mechanism for intervening on internal model behavior, allowing researchers to better understand how multimodal systems work and make them more controllable. This could have significant implications for applications such as natural language processing and computer vision. Source: https://arxiv.org/abs/2602.02060