Researchers have developed Prism, the first symbolic superoptimizer for tensor programs, achieving up to 2.2× speedup over existing superoptimizers and 4.9× speedup over compiler-based approaches. The paper was submitted to arXiv on April 16, 2026, by researchers Mengdi Wu, Xiaoyu Jiang, Oded Padon, and Zhihao Jia.
Core Innovation: Symbolic Graph Representation
Prism introduces sGraph, a symbolic, hierarchical representation that compactly encodes large classes of tensor programs by symbolically representing execution parameters. The system organizes optimization as a two-level search: it constructs symbolic graphs representing families of programs, then instantiates them into concrete implementations.
This approach enables structured pruning of provably suboptimal regions of the search space using symbolic reasoning over operator semantics, algebraic identities, and hardware constraints.
Technical Architecture Combines Rigor with Scalability
Prism's architecture includes three key components:
- Efficient symbolic graph generation for exploring program families
- Equivalence verification via e-graph rewriting to ensure correctness
- Parameter instantiation through auto-tuning for concrete implementations
These components allow Prism to bridge the rigor of exhaustive search with the scalability required for modern machine learning workloads.
Performance Results on LLM Workloads
Evaluation on five commonly used LLM workloads demonstrates:
- Up to 2.2× speedup over best existing superoptimizers
- Up to 4.9× speedup over best compiler-based approaches
- Up to 3.4× reduction in end-to-end optimization time
Addressing a Critical Gap in Tensor Optimization
Tensor program optimization is critical for LLM inference efficiency. Existing approaches face a fundamental tradeoff: exhaustive search methods are prohibitively expensive, while heuristic-based compilers produce suboptimal results. Prism's symbolic approach provides mathematical guarantees while remaining practical for production use.
The system addresses the growing need for efficient LLM inference as models continue to scale and deployment costs remain a significant concern for organizations running AI workloads.
Key Takeaways
- Prism is the first symbolic superoptimizer for tensor programs, using sGraph representation to encode program families
- Achieves 2.2× speedup over existing superoptimizers and 4.9× over compiler-based approaches on LLM workloads
- Reduces end-to-end optimization time by up to 3.4× through structured pruning of suboptimal search regions
- Bridges the gap between exhaustive search rigor and practical scalability for production ML systems
- Provides mathematical correctness guarantees through equivalence verification via e-graph rewriting