ChronoVision: Temporal Reasoning via Latent State Reconstruction

2026-08-06

AI Summary

This work addresses the limitations of multimodal large language models in visual reasoning tasks requiring multi-step temporal inference, primarily due to the inadequacy of language-based representations in capturing continuous visual dynamics. To overcome this, the authors propose ChronoVision, a framework that aligns visual logic with implicit image states through latent state reconstruction and incorporates a region-of-interest (ROI) attention mechanism to localize critical visual evidence. A composite reward function—integrating outcome correctness, process alignment, and unsupervised visual focus—is designed within a reinforcement learning paradigm guided by implicit procedural cues. The study also introduces Vbvr-VQA, a new benchmark for evaluating temporal tracking capabilities. ChronoVision achieves 74.8% (in-domain) and 71.6% (out-of-domain) accuracy on Vbvr-VQA and sets a new state-of-the-art performance of 55.0% on the cross-domain IntPhys2 benchmark.

Abstract

Multimodal large language models excel at passive perception but struggle with complex visual cognitive tasks requiring multi-step temporal reasoning. This degradation largely stems from the inherent ambiguity of language-based reasoning, which often fails to accurately articulate continuous visual transformations. To address this, we propose ChronoVision, a multimodal framework designed to align visual logic with latent imagery. During supervised fine-tuning, a Reconstructive Visual Head predicts the latent representation of the final transformed state, while an ROI Attention Locating module focuses the model on key visual evidence via semantic span queries. In post-training, we apply reinforcement learning with an implicit process grounding mechanism, guided by a composite reward function that evaluates outcome correctness, latent process alignment, and unsupervised visual focus. Furthermore, we introduce Vbvr-VQA, a novel dataset that evaluates temporal tracking by reformulating video reasoning into a strict image-ordering task. Experiments demonstrate that ChronoVision achieves state-of-the-art performance on Vbvr-VQA with 74.8% in-domain and 71.6% out-of-domain accuracy, alongside a strong 55.0% accuracy on IntPhys2, a highly challenging cross-domain benchmark.

Problem

• temporal reasoning
• multimodal large language models
• visual cognitive tasks
• latent state reconstruction
• video reasoning

Innovation

• Temporal Reasoning
• Latent State Reconstruction
• Multimodal LLM
• ROI Attention
• Reinforcement Learning