Reproduce “Real-Time Visual Attribution Streaming in Thinking Model” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Seil Kang, Woojung Han, Junhyeok Kim, Jinyeong Kim, Youngeun Kim, Seong Jae Hwang
Researchers from Yonsei University and Amazon developed VSTREAM, an amortized framework that provides real-time, object-centric visual attribution for multimodal reasoning models. This method achieves faithfulness comparable to computationally intensive baselines while significantly increasing efficiency, enabling attribution streaming at speeds up to 117 times faster than prior gradient-based approaches.
Key insights
- Semantic region unitization based on self-supervised features (DINOv3) significantly enhances attribution accuracy and interpretability over non-semantic partitions.
- A simple, data-efficient linear estimator can predict complex causal ablation effects with high fidelity, achieving strong cross-task and out-of-distribution generalization.
- Real-time attribution allows for dynamic analysis of reasoning trajectories, revealing that successful reasoning correlates with more stable visual grounding (lower tortuosity) and providing early warning signals for model failures like hallucination or reasoning errors.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.