Vision-language models typically treat the vision encoder as a fixed front end: it converts an image into features once, and then the language model does all the subsequent reasoning. The authors of AdaLoop argue this is a limitation, because different questions may require different visual details or levels of abstraction.
AdaLoop proposes to make the vision encoder "think" by looping over the image representation in a question-aware manner. Instead of a single forward pass, the encoder iteratively refines its features, guided by the text query. This allows visual reasoning to happen earlier in the pipeline.
The paper's abstract is brief, so details on architecture and results are not yet available. The core idea, however, is a clear departure from the standard static-encoder design, suggesting that visual perception itself can be a dynamic, question-driven process.