ChronoVision: Temporal Reasoning via Latent State Reconstruction
ChronoVision is a multimodal framework for temporal visual reasoning that supplements language-based reasoning with latent visual-state reconstruction. During supervised fine-tuning, a Reconstructive Visual Head predicts the latent representation of the final transformed state, while an ROI Attention Locating module identifies relevant visual evidence through semantic span queries. Post-training uses reinforcement learning with rewards for answer correctness, latent process alignment, and unsupervised visual focus. The authors introduce Vbvr-VQA, an image-ordering benchmark, and report 74.8% in-domain, 71.6% out-of-domain, and 55.0% on IntPhys2 accuracy.
Why it's worth reading
The paper is timely because it tests whether explicit latent-state reconstruction can reduce multimodal models’ weakness on multi-step visual transformations, including out-of-domain and cross-benchmark evaluation.