NANOZK: Layerwise Zero-Knowledge Proofs for Verifiable Large Language Model Inference

Researchers led by Zhaohui Geoffrey Wang have developed NANOZK, a breakthrough zero-knowledge proof system called METHOD that solves a critical trust problem in AI services. When users query proprietary LLM APIs from companies like OpenAI or Anthropic, they currently have no way to verify that the expensive frontier model they're paying for was actually used to generate responses. Providers could secretly substitute cheaper models, apply aggressive quantization, or return cached outputs—all while charging premium prices. METHOD provides cryptographic assurance that outputs correspond to the computation of a specific model architecture.

The system's key innovation is a layerwise proof framework that exploits transformer architecture's natural decomposition into independent layers. Each layer generates a constant-size proof of just 5.5KB (2.1KB for attention + 3.5KB for MLP) regardless of model width, with verification taking only 24 milliseconds. This approach sidesteps scalability barriers that made monolithic verification impractical for large models. The team developed lookup table approximations for non-arithmetic operations like softmax and GELU that introduce zero measurable accuracy loss, preserving model perplexity exactly while enabling efficient cryptographic proofs.

Compared to existing systems like EZKL, METHOD achieves dramatic improvements: 70x smaller proofs and 5.7x faster proving time for models with dimension d=128, while maintaining formal soundness guarantees (epsilon < 1e-37). The system also introduces Fisher information-guided verification for scenarios where proving all layers is impractical, providing flexible security-performance tradeoffs. This represents the first practical solution for verifiable LLM inference at scale, potentially transforming how enterprises audit AI service providers and ensuring they receive what they pay for in mission-critical applications.