QUIVER: A Formal Framework for Quantifying Perturbation Propagation and Bifurcation in Compound AI Systems
Researchers have introduced QUIVER, a new formal framework designed to measure how perturbations propagate through compound AI systems that chain multiple large language model calls. This framework quantifies sensitivity by classifying edges as amplifiers or absorbers and decomposes variation into value drift, structural path divergence, and iteration count divergence. QUIVER also identifies bifurcation thresholds, the minimum perturbation causing structural execution path changes, and assesses distribution faithfulness. Validation on enterprise and public pipelines demonstrated QUIVER's ability to reveal distinct sensitivity profiles and predict nodes prone to trajectory bifurcation. AI
IMPACT Provides a method to analyze and improve the robustness of complex, multi-stage AI systems.