University of Waterloo

Researchers at the University of Waterloo developed zkLLM, a zero-knowledge proof system for large language model inference S-0023. Their verification work includes:

• Their CCS 2024 paper combines lookup arguments for non-arithmetic operations with a dedicated proof for attention S-0023. Its zkLLM benchmark proves one forward pass of LLaMa-2-13B over a 2,048-token input in 803 seconds on an NVIDIA A100, with a 188 kB proof and 3.95-second verification S-0023. See Zero-knowledge proofs of inference.
• The authors released the zkLLM code as an artifact after the conference's artifact evaluation, archived on Zenodo S-1108. The GitHub repository was made read-only in July 2025, and its author states that the project is no longer actively maintained S-1108.
• The authors wrote that zero-knowledge proofs of LLM training "may pose insurmountable challenges" S-0023; see Zero-knowledge proofs of training constraints.