Problem
Research questions and friction points this paper is trying to address.
distributed parallel AI programs
non-deterministic execution
reliability gaps
high-dimensional inputs
non-linear operations
Directed Neuro-Symbolic Stochastic Execution for Verification of Distributed Parallel AI Programs
🤖 AI Summary
This work addresses the challenge of effectively verifying distributed parallel AI programs, whose non-deterministic execution, high-dimensional inputs, and nonlinear operations render traditional testing inadequate. The authors propose DNSSE, a hybrid testing framework that uniquely integrates large language model–guided schedule prediction, symbolic constraint solving, and coverage-guided random mutation. The approach models programs as non-deterministic transition systems and formalizes correctness properties using linear temporal logic. Theoretical analysis establishes the bounded completeness and probabilistic completeness of the hybrid solver. Empirical evaluation on PyTorch and Ray demonstrates that DNSSE detects 2.9% more concurrency bugs than the strongest baseline and improves average branch coverage from 68.6% to 91.6% across five real-world distributed AI benchmarks.
📝 Abstract
Distributed parallel Artificial Intelligence (AI) programs expose reliability gaps that conventional testing cannot close: parallel executions are non-deterministic, and AI workloads bring high-dimensional inputs and non-linear operations that defeat fuzzing and symbolic execution in isolation. We present Directed Neuro-Symbolic Stochastic Execution (DNSSE), a hybrid testing framework that couples schedule prediction guided by a Large Language Model (LLM) with symbolic constraint solving and coverage-guided stochastic mutation. We model distributed AI executions as non-deterministic transition systems, specify correctness in linear temporal logic, and prove soundness, bounded completeness, and probabilistic completeness of the hybrid solver, together with an expected-cost analysis of LLM-guided schedule exploration. A scalable implementation on PyTorch and Ray detects 2.9% more concurrency bugs than the strongest baseline and raises average branch coverage from 68.6 % to 91.6 % across five realistic distributed AI benchmarks.
Innovation
Methods, ideas, or system contributions that make the work stand out.
Neuro-Symbolic Execution
LLM-Guided Scheduling
Distributed AI Verification
Stochastic Mutation
Temporal Logic Specification
🔎 Similar Papers
No similar papers found.
