ReasonX: Declarative Reasoning on Explanations

Explanations can be made more interactive and user-friendly: ReasonX allows people to ask questions about ML decisions using a declarative query system, moving beyond static explanations to enable users to explore "what-if" scenarios and understand decisions at different levels of detail—from specific examples to more general patterns.

Background knowledge should be integrated into explanations: The tool lets users incorporate their own common sense and domain expertise as constraints when examining model decisions, making explanations more relevant to real-world context rather than treating ML models as isolated systems.

Decision trees can serve as bridges to understand complex models: ReasonX works with decision trees as either direct models or as "surrogate models" that approximate black-box AI systems, suggesting that simpler, interpretable models can be valuable tools for explaining more opaque AI systems to non-technical stakeholders.

Reasoning about alternatives improves understanding: By using mathematical optimization techniques to explore both "factual" explanations (why the model made this decision) and "contrastive" ones (why it didn't make an alternative decision), ReasonX provides more complete explanations that help users understand the boundaries of model behavior.

Multiple levels of abstraction matter for different audiences: The tool's ability to reason from fully specified examples to partially constrained scenarios means different stakeholders—from domain experts to general users—can engage with explanations at the appropriate level of technical detail for their needs.