Prism: First Symbolic Superoptimizer for Tensor Programs Enables Structured Pruning of Suboptimal Implementations

arXiv·low signal

Prism (arXiv 2604.15272) introduces the first symbolic superoptimizer for tensor programs using sGraph, a hierarchical representation that compactly encodes large program families by symbolically representing execution parameters. The two-level search constructs symbolic graphs representing program families, then instantiates concrete implementations — enabling structured pruning of provably suboptimal candidates. For ML engineers optimizing inference pipelines, symbolic superoptimization could automate the manual kernel tuning that currently requires deep hardware expertise.