SMART: Symbolic ML Performance Modeling via Design Doc Regeneration
September 7, 2026
SMART uses a DAG of natural-language design documents to drive code generation for ML performance modeling. Instead of manual patching, coding sub-agents regenerate the entire implementation from these docs to prevent technical debt in rapidly evolving ML systems.
HOW THIS AFFECTS YOU
●
builderYou can maintain performance libraries by updating documentation rather than refactoring complex codebases.
●
researcherThis introduces a new paradigm for maintaining highly specialized software through LLM-driven regeneration.