DNSSE: LLM-Guided Testing Framework for Distributed Parallel AI Programs
A recent submission to arXiv (2608.07947) presents Directed Neuro-Symbolic Stochastic Execution (DNSSE), a novel framework designed for testing distributed parallel AI applications. This hybrid approach integrates LLM-assisted schedule forecasting with symbolic constraint resolution and coverage-driven stochastic mutation to tackle reliability issues in non-deterministic parallel processes. DNSSE conceptualizes distributed AI operations as non-deterministic transition systems and defines correctness using linear temporal logic. The authors demonstrate the soundness, bounded completeness, and probabilistic completeness of their hybrid solver, alongside an analysis of expected costs related to LLM-guided schedule exploration. Implemented on PyTorch and Ray, the system identifies 2.9% more concurrency bugs than the leading baseline while improving average coverage. The paper is accessible via the provided URL.
Key facts
- Paper arXiv:2608.07947 introduces DNSSE, a hybrid testing framework for distributed parallel AI programs.
- DNSSE combines LLM-guided schedule prediction with symbolic constraint solving and coverage-guided stochastic mutation.
- The framework models distributed AI executions as non-deterministic transition systems.
- Correctness is specified in linear temporal logic.
- The hybrid solver is proven to be sound, bounded complete, and probabilistically complete.
- An expected-cost analysis of LLM-guided schedule exploration is provided.
- Implementation on PyTorch and Ray detects 2.9% more concurrency bugs than the strongest baseline.
- The paper is announced as a new submission on arXiv.
Entities
Institutions
- arXiv
- PyTorch
- Ray