Papers · Proving inference · NANOZK

NANOZK

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

Zhaohui Geoffrey Wang

Venue
arXiv:2603.18046
Date
2026-03-17
Proof system
Layerwise proofs, constant-size per layer regardless of model width
Numbers from
primary
Paper
https://arxiv.org/abs/2603.18046
PDF
https://arxiv.org/pdf/2603.18046
Quantization
not stated

Reported benchmarks

modelproving time sproof size mbverification time sbaseline comparison
Transformer block (d=128)430.00550.024vs EZKL: 70x smaller proofs, 5.7x faster proving.

proof_size_note: 5.5 KB per layer = 2.1 KB attention + 3.5 KB MLP

Notes

Threat model is model substitution: a provider silently serving a cheaper or more aggressively quantized model.

Quantization, in full

bits
None
scheme
Lookup-table approximations for non-arithmetic operations
accuracy_retention
claims lookup approximations "preserve model perplexity exactly"
note
An 'exactly preserves perplexity' claim is extraordinary and should be treated as unverified.

Our reading

Citation neighbourhood

Builds on
  • none recorded
Cited by, in this corpus
  • none recorded

Edges are a proxy: paper A's text mentions B anywhere (body or bibliography). See the full graph.

Discussed in

System by systemThe landscapeProof systemsQuantization, the hidden variableWhat is actually being proven
Other recorded fields
quality_flag: 'CAUTION: single-author preprint, not peer reviewed. The published abstract
  literally contains an unsubstituted "METHOD" placeholder where the system name should
  be, which suggests it was not carefully proofread. Verify every claim before citing.'