ZonoGPT: Towards An Abstract Domain for Verifying Large GPT Models

2026-09-28

AI Summary

This study addresses the challenge of balancing verification depth and precision for large Transformer models, which existing methods struggle to achieve. We propose ZonoGPT, a novel abstract domain that leverages structured zonotopes and generator reduction mechanisms to ensure space complexity remains independent of network depth. Furthermore, this work pioneers the verification of large-scale Transformers with standard architectures by introducing block-level fused transformations and affine handling of GELU activations to effectively preserve feature correlations. The proposed approach successfully verifies the 300-million-parameter GPT-2 Medium model across 1,339 instances spanning both textual and visual tasks. These results represent a significant breakthrough in achieving efficient and precise formal verification for large-scale Transformer architectures.

Abstract

Transformer-based models are widely used for reasoning, coding, and multimodal agentic tasks. To provide formal assurance of desirable behaviors, such as robustness, safety, and fairness, neural network verification techniques prove required properties and provide auditable guarantees before deployment. However, prior work remains limited to small or restricted Transformers, and maintaining precision across deep models remains challenging. In this work, we introduce ZonoGPT, an abstract domain for verifying large transformers that maintains a space complexity independent of network depth. ZonoGPT uses a structured zonotope and a generator reduction mechanism to efficiently preserve correlations. To maintain precision, it introduces block-specific fused transformations for Attention and LayerNorm that retain feature relations, along with an affine transform for GELU that preserves generator relations. These mechanisms enable \tool{} to be the first approach to verify standard architectures, scaling to official HuggingFace models up to GPT-2 Medium (24 blocks, 300M+ parameters) and successfully verifying 1,339 instances across text and vision tasks.

Problem

• Transformer verification
• Large language models
• Formal verification
• Abstract domain
• Robustness

Innovation

• Abstract Domain
• Zonotope
• Transformer Verification
• Generator Reduction
• Fused Transformations