Enhancing Reliability Assurance for DNN against Numerical Defect with Large Language Models

Deep Neural Networks (DNNs) are increasingly deployed in safety-critical systems, yet their numerical reliability remains a pressing concern due to defects like NaN/INF occurrences during training/inference. Existing static analyzers detect operator-level risks but lack the ability to establish contextual mappings between operator-level anomalies and their originating code implementations, thereby limiting their practical applicability in real-world debugging scenarios. To address this limitation, this paper introduces a novel hybrid method that synergizes risk sensitive paths derived from static analysis with Large Language Model(LLM)-guided semantic reasoning. By anchoring LLMs to identified risk-sensitive computation paths, the proposed method precisely correlates operator-level numerical defects with code-level vulnerabilities, while mitigating LLM hallucinations. Evaluated on a widely-used real-world benchmark, the proposed method achieves 100% defect detection accuracy, matching state-of-the-art static analyzers, but significantly outperforms them in code-level root-cause localization. This work pioneers a new paradigm for DNN reliability assurance by combining the static numerical analysis with the semantic reasoning power of LLMs, establishing a practical numerical defect analysis pipeline.