Reproduce “Entropogram-guided Visual Realignment Decoding: Capturing the Rhythm of Multimodal Reasoning” 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
Jiaqi Deng, Zhe Fu, Yakun Chen, Dichang Zhang, Zonghan Wu, Xiaoshui Huang, Huan Huo, Guandong Xu
SEED is a training-free decoding method that monitors entropy during multimodal reasoning and injects visually guided representations at potentially unreliable states. On Qwen3-VL-8B-Instruct, it raises accuracy from 78.21% to 79.97% on M³CoT and improves results in four of five reasoning settings. The entropy-hallucination relationship is correlational, and outcomes vary across tasks and backbones.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.