Searcharxiv⌕ Search

arXiv subjects

Qiuhan Gu

Publications and source records attributed to Qiuhan Gu.

2 recordsLinked to original sources

A Nonlinear Abstract Domain for Bayesian Neural Networks

Bayesian neural networks (BNNs) explicitly model epistemic uncertainty by leveraging probability distributions over network parameters, making them better suited for risk-aware decision-making. Yet soundly propagating uncertainty through the network and capturing the entire output distribution to support predictive uncertainty quantification remains a fundamental challenge. Existing approaches are either sampling-based, providing no formal guarantees, or propagate only selected statistical properties, such as expectations and entropy, limiting their applicability to general probabilistic queries. To address these challenges, we present the first nonlinear abstract domain for soundly approximating complete probability distributions of BNNs. The abstract domain builds upon a novel marginal collecting concrete semantics that soundly abstracts the classical joint-measure semantics. It combines support partitioning with exponential-affine bounds to represent probability measures, capturing both discrete and continuous components of mixed distributions as well as nonlinear density shapes over bounded and unbounded supports. We develop sound abstract transformers that can efficiently handle high-dimensional affine transformations and nonlinear ReLU activations, while remaining highly parallelizable. We further extend the framework beyond BNNs to support bounded-loop probabilistic programs with Bayesian inference to demonstrate its generality. We implement the resulting abstract interpreter as a GPU-accelerated analyzer. Experimental results show that the framework significantly improves precision over existing probabilistic reasoning approaches while soundly capturing the complete output distribution, demonstrating strong efficiency and scalability.

cs.PL↗

SAIL: Sound Abstract Interpreters with LLMs

How to construct globally sound abstract interpreters to safely approximate program behaviors remains a bottleneck in abstract interpretation. In this paper, we show the potential of using state-of-the-art LLMs to automate this tedious process. Focusing on the neural network verification area, we synthesize non-trivial sound abstract transformers across diverse abstract domains using LLMs to search within infinite space from scratch. We formalize the synthesis task as a constrained optimization problem, for which we design a novel mathematically grounded cost function that measures the degree of unsoundness of each generated candidate transformer, while enforcing hard syntactic and semantic validity constraints. Building on this formulation, we introduce SAIL, a novel unified framework that combines model generation, syntactic and semantic validation, and cost-function-based refinement to synthesize globally sound abstract transformers. Evaluation results show that SAIL not only matches the performance of manually designed transformers, but also is able to synthesize sound and high-precision transformers that do not exist in the literature for complex non-linear operators.

cs.PL↗