In many human-AI systems, the AI does not directly execute an action. Instead, a candidate action is first evaluated by another decision-making entity—often a human—under a potentially different objective or decision model. The new arXiv paper "From Proposal to Execution: Modeling Interaction Semantics in Shared Autonomy" highlights this as a common but under-studied aspect of shared autonomy.
The authors note that standard decision-theoretic frameworks do not fully account for this interaction step. By modeling the semantics of how proposals are passed, interpreted, and evaluated, the work aims to better capture the full loop from AI proposal to final execution.
The paper's contribution is conceptual: it shifts attention from isolated decision-making to the interaction between proposer and evaluator. This matters for designing systems where human oversight is part of the control loop, not an afterthought.