Provenance Tracking in AI Compilers through the Lens of Coalgebra

Researchers present a coalgebra-based approach to tracking tensor and operator provenance through AI compiler transformations, addressing the challenge of maintaining computational lineage during aggressive graph rewrites. The method uses observational semantics rather than identifier propagation, with a prototype implementation called COVAN demonstrating practical viability with minimal engineering overhead.

AI compilers face a fundamental problem: as they optimize computation graphs through normalization and lowering passes, the original provenance of tensors and operators becomes obscured or lost entirely. This matters because developers need reliable provenance for debugging, validating transformations, and applying platform-specific optimizations—yet current solutions require invasive modifications to compiler infrastructure or produce unreliable results when graph rewrites eliminate intermediate nodes.

The research addresses this by shifting from traditional identifier-tracking approaches to observational semantics grounded in coalgebraic theory. Instead of propagating IDs through compilation stages, the method observes actual computational transformations and reasons about provenance through what can be empirically verified in the compiled graph. This theoretical foundation provides mathematical guarantees about bisimulation, ensuring provenance relationships survive even aggressive node elimination.

For the AI and machine learning infrastructure sector, this work represents a meaningful advance in compiler reliability. Better provenance tracking enables more confident optimization strategies, faster debugging cycles for compiler developers, and improved transparency for systems running on heterogeneous hardware where platform-specific postprocessing is essential. The COVAN prototype demonstrates this isn't purely theoretical—the approach scales to real compilation pipelines without significant engineering burden.

The practical implications extend to any organization using AI compilers for production inference or training. As models grow larger and hardware becomes more specialized, the ability to reliably trace computational origins becomes increasingly critical for performance validation and regulatory compliance in sensitive domains. The lightweight nature of this approach suggests potential adoption in existing compiler frameworks.

• →Coalgebra-based provenance tracking preserves tensor lineage even when intermediate compiler nodes are eliminated during optimization
• →Observational semantics approach requires minimal invasive changes to existing AI compiler infrastructure
• →COVAN prototype demonstrates practical implementation with stable provenance across full compilation pipelines
• →Better provenance tracking enables improved debugging, validation, and platform-specific optimization in AI compilers
• →Method addresses critical gap between aggressive graph rewrites and developer need for reliable computational lineage