Overview
In many human-AI systems, candidate actions are not directly executed. Instead, an action proposed by one entity is first evaluated by another decision-making entity that may operate under a different objective or decision model. Researchers have developed a semantics-aware decision framework that models this multi-stage interaction explicitly. By introducing a formal intermediate layer between action proposal and physical execution, the framework separates context-dependent evaluation from transition dynamics while retaining full compatibility with standard Markov decision process (MDP) planning algorithms.
Research Context
Standard decision-theoretic models define transition dynamics directly across state-action pairs. Under this convention, the evaluation step between interacting entities is absorbed entirely into the environment's transition model rather than represented as a distinct operational phase. While mathematically compact, absorbing intermediate evaluation into state transitions obscures the underlying interaction semantics and prevents the modular specification of how proposals are evaluated across different interaction contexts.
Approach
The proposed framework formalizes decision-making through a three-stage sequence: a candidate action is proposed, evaluated under an explicit interaction context, and subsequently resolved into an executed outcome. This architecture relies on two mathematical components:
- Joint Action Language: Represents the proposals and corresponding interactive inputs from participating entities.
- Interpretation Function: Maps each joint action and interaction context directly to a semantic outcome.
Within this formulation, transition and reward dynamics are induced through the resulting semantic outcomes rather than directly from raw state-action pairs. This establishes an explicit representational layer for interaction semantics. Because semantic outcomes serve as modular, organized units for interaction principles, application-specific evaluation mechanisms can be configured and reused across distinct interaction instances. Crucially, this separation provides a structured interface to specify evaluation logic without requiring structural modifications to the underlying planning algorithm.
Findings
The study demonstrates that introducing explicit semantic outcomes allows interaction patterns—such as advisory modes and authorization mechanisms—to be represented directly within shared autonomy settings. Evaluating candidate actions through an interpretation function illustrates precisely how interaction context alters final decision-making. Furthermore, because transition and reward functions are induced via these resolved semantic outcomes, the resulting model remains fully solvable using standard MDP planning methods.
Potential Applications
The framework is specifically instantiated in shared human-AI autonomy, providing formal representations for interactive configurations where an action must pass through advisory evaluation or explicit authorization before physical execution.