Do Vision-Language-Action Models truly follow your language instructions? We present When Vision Overrides Language: Evaluating and Mitigating Counterfactual Failures in VLAs. They promise to ground language instructions in robot control, yet in practice, often fail to follow language faithfully. 📄 Paper: arxiv.org/abs/2602.17659 🌐 Project: vla-va.github.io/ 💡 Highlights Vision shortcuts and counterfactual failures. When given instructions that lack strong scene-specific supervision, they default to well-learned scene-specific behaviors regardless of language intent. Counterfactual benchmark. We introduce LIBERO-CF, the first counterfactual benchmark for evaluating language following in VLAs. Our evaluation reveals that counterfactual failures are prevalent yet underexplored across state-of-the-art VLAs. Our solution. We propose Counterfactual Action Guidance (CAG), a simple plug-and-play dual-branch inference scheme that strengthens language conditioning without changing pretrained VLA architectures or weights. Experiments. CAG is effective across multiple dimensions of language grounding, consistently improving both language grounding and task success on under-observed tasks. #VLA #Robotics #Vision #Language

🤖Real-world Experiments We study different aspects of language grounding. Each scene is designed with three possible tasks: one well-learned in-domain task with sufficient demonstrations, and two under-observed tasks defined by counterfactual instructions. Across object recognition, spatial reasoning, goal targeting, out-of-distribution generalization, and long-horizon reasoning, our proposed CAG consistently improves the performance of pi0.5 and reduces counterfactual failures. [5/6]