DiffTilt Improves Falsification of Safety-Critical Systems

Researchers have developed DiffTilt, a novel distributional framework designed to improve the falsification of safety-critical systems. This method addresses the challenge of discovering rare failures in autonomous and cyber-physical systems, where existing approaches suffer from "multiplicative rarity effects"—the simultaneous scarcity of both failure-inducing inputs and execution traces. DiffTilt employs exponential tilting on a diffusion model-induced joint distribution of environments and executions. The framework demonstrates that diffusion-guided sampling can be interpreted as importance sampling, where guidance scores optimally reallocate probability mass towards failure-relevant behaviors. This tilting provably amplifies failure probability and consistently outperforms conditional sampling, which is limited by the multiplicative rarity problem. A key advantage is that the joint generative model acts as a reusable prior for scenarios, and expensive system simulations are only used to learn a scoring function for scenario quality, allowing for their selective and adaptive use. Evaluated on ARCH-COMP benchmarks and a new tractor-trailer benchmark, DiffTilt achieves competitive or superior falsification performance compared to state-of-the-art methods, especially when specifications are well-defined.