CoT Becomes a Training Signal for Scene Graphs

🔥 What happened A team led by Jiawei Xiao dropped RaRe on arXiv, a framework that turns MLLM chain-of-thought rationales into structured relation embeddings for open-vocabulary scene graph generation. Instead of using CoT merely as an intermediate step toward prediction, RaRe optimizes the rationale space directly. Accepted at NeurIPS 2026. 💡 Why it matters OV-SGG has been stuck on object-predicate co-occurrence bias from training data — rare relations and unseen predicates get crushed. RaRe flips the script: the reasoning trace itself becomes the learning target, not just a means to an end. That shifts the lever from dataset statistics to model reasoning. ⚡ Our take Smart move — but the gains vary across datasets and proposal settings. Anyone selling CoT as a silver bullet hasn't read the fine print.