Paper 2026/2907

Sparse-zkVC: Sparsity-Aware R1CS Synthesis for Zero-Knowledge Proofs of Sparse LLM Inference

yanzhe cheng, Southwest University
Luxing Yang, Deakin University
Wei Ren, Southwest University
Abstract

Zero-knowledge proofs for machine-learning inference are often limited by dense arithmetization: even after model pruning, R1CS generators may instantiate witness variables and linear-combination terms according to the original tensor shape. Sparse-zkVC extends zkVC's CRPC+PSQ matrix-multiplication generator with mask-guided synthesis. Given a public static pruning mask, it allocates variables only for retained weights and active prefix-sum states while advancing CRPC coefficients over the original dense index order. This preserves coefficient alignment, keeps the constraint count at the inner dimension $d$, and reduces the dominant variable term from $h d m$ to $h\cdot NNZ$, where $NNZ$ counts retained weights. The prototype removes unused variables and linear-combination terms during frontend synthesis. On the zkVC-aligned $128\times64\times49$ benchmark, 99\% sparsity reduces variables from 412,736 to 12,191 and improves online proving time by 20.2$\times$, with all generated proofs verifying successfully. Long-context and LLaMA-7B linear-operator stress tests characterize the resulting efficiency gains and remaining single-node memory limits.

Metadata
Available format(s)
PDF
Category
Applications
Publication info
Preprint.
Keywords
Zero-knowledge proofsverifiable computationZKMLLLM pruningsparsity-aware R1CS synthesisMSM
Contact author(s)
orangeeeeee @ email swu edu cn
History
2026-11-21: approved
2026-11-16: received
See all versions
Short URL
https://ia.cr/2026/2907
License
Creative Commons Attribution
CC BY

BibTeX

@misc{cryptoeprint:2026/2907,
      author = {yanzhe cheng and Luxing Yang and Wei Ren},
      title = {Sparse-{zkVC}: Sparsity-Aware {R1CS} Synthesis for Zero-Knowledge Proofs of Sparse {LLM} Inference},
      howpublished = {Cryptology {ePrint} Archive, Paper 2026/2907},
      year = {2026},
      url = {https://eprint.iacr.org/2026/2907}
}
Note: In order to protect the privacy of readers, eprint.iacr.org does not use cookies or embedded third party content.