Latent visual reasoning aims to achieve multimodal reasoning via hidden‑state computation instead of explicit textual chains of thought. Yet embedding visual information in a latent state does not guarantee that the model actually uses that state when answering, especially when alternative image‑conditioned routes remain available. We introduce Causal Visual Recurrent Reasoning (CVRR), which preserves pretrained visual‑language competence while forcing recurrent computation to become the sole image‑conditioned path. CVRR starts recurrence from the question hidden state after the pretrained model has already incorporated the image, then repeatedly updates this state while re‑reading the same fixed visual evidence. Before decoding, all visual states and the original multimodal KV cache are removed, leaving only the final recurrent state to carry image‑conditioned information to the answer. Across the $V^*$, MMVP, BLINK, and MME‑RealWorld‑Lite benchmarks, CVRR retains strong performance under this strict interface, whereas compatible latent reasoners retrained under the same constraint fail to recover comparable visual competence. Causal interventions further show that predictions remain sensitive to recurrent content when the question is fixed, and that persistent visual evidence causally revises the recurrent trajectory. These findings separate latent informativeness from latent computation actually used for prediction.
Review