// ABSTRACT
Users querying proprietary LLM APIs lack cryptographic verification that the claimed model was actually used. NANOZK enables verifiable inference by decomposing transformer computations into independent layer operations, generating constant-size proofs regardless of model width. Lookup table approximations preserve model accuracy (0% perplexity degradation across 21 configurations), while Fisher information-guided verification enables practical budget allocation. Achieves 70x smaller proofs and 5.7x faster proving than EZKL, with formal soundness guarantees (epsilon < 1e-37). LLaMA-3-1B: 620ms proof generation, 2.4KB proof size, 22ms verification on CPU.
// BIBTEX
@inproceedings{wang2026nanozk,
title = {NANOZK: Layerwise Zero-Knowledge Proofs for Verifiable Large Language Model Inference},
author = {Zhaohui Geoffrey Wang},
booktitle = {VerifAI Workshop at International Conference on Learning Representations (ICLR)},
year = {2026},
month = {3},
eprint = {2603.18046},
archivePrefix = {arXiv},
}