Real-Time Visual Attribution Streaming in Thinking Model

We present an amortized framework for real-time visual attribution streaming in multimodal thinking models. When these models generate code from a screenshot or solve math problems from images, their long reasoning traces should be grounded in visual evidence. However, verifying this reliance is challenging: faithful causal methods require costly repeated backward passes or perturbations, while raw attention maps offer instant access, they lack causal validity. To resolve this, we introduce an amortized approach that learns to estimate the causal effects of semantic regions directly from the rich signals encoded in attention features. Across five diverse benchmarks and four thinking models, our approach achieves faithfulness comparable to exhaustive causal methods while enabling visual attribution streaming, where users observe grounding evidence as the model reasons, not after. Our results demonstrate that real-time, faithful attribution in multimodal thinking models is achievable through lightweight learning, not brute-force computation.

Large vision-language models can now solve problems by looking at images and writing long step-by-step reasoning traces. However, it is often hard to tell whether each step of the reasoning is actually based on the right part of the image or is just a plausible-sounding explanation. This paper introduces a method that shows, in real time, which image regions most influence the model as it reasons. Instead of running slow and expensive checks after the model finishes, the method learns a lightweight predictor that estimates reliable visual attributions during generation. This helps users inspect whether a multimodal reasoning model is using relevant visual evidence, which can make debugging and auditing these systems more practical.