@misc{indiciaeb8f496c41677, title = {When Does a Video-Language Model Stop Watching? Reward Strength Controls the Formation and Reversal of Visual Shortcuts in Multimodal RLVR}, author = {Zekun Xu}, year = {2026}, url = {https://arxiv.org/abs/2606.22043}, note = {Source identifier: 2606.22043} }