Abstract
Real-world data often appears in diverse, disjoint forms -- with varyingschemas, inconsistent semantics, and no fixed feature ordering -- making itchallenging to build general-purpose models that can leverage informationacross datasets. We introduce ASPIRE, Arbitrary Set-based Permutation-InvariantReasoning Engine, a Universal Neural Inference model for semantic reasoning andprediction over heterogeneous structured data. ASPIRE combines apermutation-invariant, set-based Transformer with a semantic grounding modulethat incorporates natural language descriptions, dataset metadata, andin-context examples to learn cross-dataset feature dependencies. Thisarchitecture allows ASPIRE to ingest arbitrary sets of feature--value pairs andsupport examples, align semantics across disjoint tables, and make predictionsfor any specified target. Once trained, ASPIRE generalizes to new inferencetasks without additional tuning. In addition to delivering strong resultsacross diverse benchmarks, ASPIRE naturally supports cost-aware active featureacquisition in an open-world setting, selecting informative features undertest-time budget constraints for an arbitrary unseen dataset. Thesecapabilities position ASPIRE as a step toward truly universal, semantics-awareinference over structured data.