🔁 CVRR — Reason Through the Latent
Causal Visual Recurrent Reasoning on top of Qwen2.5-VL-7B-Instruct. The lower
decoder is frozen up to the causal visual-read boundary; a single decoder layer is then
reused as a shared recurrent transition that refines the question state while keeping
the native visual rows as fixed evidence. Only the final recurrent state reaches the
answer decoder — the original visual rows and multimodal prefix cache are cut off, which
makes the latent visual reasoning causally necessary rather than merely present.
Ask a question about an image — open-ended, or in the benchmark-style multiple-choice format the model was tuned on. Decoding is deterministic (greedy) by design. The recurrent transition is supervised on short VQA answers, so crisp, targeted questions work best; long free-form descriptions drift.