Consistency Amplifies: How Behavioral Variance Shapes Agent Accuracy

This arXiv collection highlights security-relevant advances and risks in LLM-driven software and hardware engineering. Key findings: (1) "Consistency Amplifies" shows that low behavioral variance correlates with higher accuracy but often amplifies consistently wrong interpretations (e.g., 71% of one model's failures were consistent wrong assumptions). (2) "Search-Induced Issues (SII)" identifies a novel failure mode where web-augmented LLMs are misled by unreliable or malicious pages, causing incorrect (and potentially insecure) code. (3) A large-scale empirical study of Java library exploits: