ChronoVision: Temporal Reasoning via Latent State Reconstruction
ChronoVision trains a multimodal model to reason through visual change by reconstructing the latent final state, not just describing steps in text.
The paper says language-based reasoning can lose precision when visual transformations are continuous or multi-step. Its framework adds a reconstructive visual head for the expected final state and an ROI attention module to focus on key evidence. Post-training uses reinforcement learning rewards for answer correctness, latent process alignment, and visual focus. The authors also introduce Vbvr-VQA, an image-ordering benchmark for temporal tracking, where ChronoVision reports 74.8% in-domain and 71.6% out-of-domain accuracy. HF Daily Papers' note
The paper says language-based reasoning can lose precision when visual transformations are continuous or multi-step. Its framework adds a reconstructive visual head for the expected final state and an ROI attention module to focus on key evidence. Post-training uses reinforcement learning rewards for answer correctness, latent process alignment, and visual focus. The authors also introduce Vbvr-VQA, an image-ordering benchmark for temporal tracking, where ChronoVision reports 74.8% in-domain and 71.6% out-of-domain accuracy. HF Daily Papers' note
score 4