Researchers have proposed Causal Visual Recurrent Reasoning (CVRR), a method designed to test whether multimodal models genuinely rely on latent visual computation when answering questions, rather than merely carrying visual information in hidden states while using other image-conditioned pathways.
The core problem the paper identifies is that the presence of visual information in a latent state does not prove the model uses that state for prediction, especially when alternative image-conditioned routes remain available. CVRR addresses this by preserving pretrained visual competence while making recurrent computation the required image-conditioned path to the answer.
Mechanically, CVRR initializes recurrence from the question hidden state after a pretrained vision-language model has already incorporated the image. It then repeatedly updates this state while re-reading the same fixed visual evidence. Before decoding, the visual states and the original multimodal KV cache are removed, so only the final recurrent state carries image-conditioned information into the answer.
According to the paper, CVRR retains strong performance under this strict interface across the V*, MMVP, BLINK, and MME-RealWorld-Lite benchmarks. By contrast, compatible latent reasoners fail to recover comparable visual competence even when retrained under the same constraint.
Causal interventions reported in the work further show that predictions remain sensitive to recurrent content when the question is held fixed, and that persistent visual evidence causally revises the recurrent trajectory. The authors frame these results as distinguishing latent informativeness from latent computation that is actually used for prediction.