Twin Worlds Framework for Equivariance-Based Reasoning Abstention

Twin Worlds improves grounding in knowledge-intensive tasks by using equivariance-based abstention to detect when entity mentions trigger ungrounded memorized associations. The framework forces models to abstain when reasoning steps do not transform correspondingly under entity changes.