Backlog

Backlog

Indexed but not yet read. Honest about what we have not done.

Known of, not read

These have no entry in the corpus proper — no benchmark numbers, no provenance, no page. They are listed so that the size of what we have not done stays visible.

PaperDateCategoryNote
Verifiable Fine-Tuning for LLMs: Zero-Knowledge Training Proofs Bound to Data Provenance and Policy2025-10training
VeriLLM: A Lightweight Framework for Publicly Verifiable Decentralized Inference2025-09inference
ZKProphet: Understanding Performance of Zero-Knowledge Proofs on GPUs2025-09infrastructure
Range-Arithmetic: Verifiable Deep Learning Inference on an Untrusted Party2025-05inferenceTurns rounding-after-matmul and ReLU into arithmetic steps verifiable by sum-check + concatenated range proofs; avoids Boolean encoding and large lookup tables. Real zkML inference optimization, fully sound -- not a sampling scheme.
Engineering Trustworthy Machine-Learning Operations with Zero-Knowledge Proofs2025-05infrastructure
Bionetta: Efficient Client-Side Zero-Knowledge Machine Learning Proving2025inferenceCited by both DeepProve [ea25a] and Jolt Atlas. Groth16/UltraGroth, client-side, EVM on-chain verification; weights hardcoded into the circuit (0-constraint fixed-matrix mul). General ML, NOT LLM-specific, no general ONNX support. Not a graph candidate; track for completeness.
Zhan et al.2024-11inference
Lookup arguments for decision trees (cq+, zkcq+, cq++)2024-04testing
South et al.2024-02inference
MSCProof2023-10inference
Fan et al.2023-05inference
ezDPS2022-12inferenceZero-knowledge data processing pipeline (not just the model).
Kang et al. (Scaling up Trustless DNN Inference)2022-10inference
Singh et al. (decision trees)2022-10inference170-200s proving, 0.4s verify.
Hydra2021-12inference
Hydra2021-12trainingInteractive quantization algorithm: rounds+freezes bottom layers, then quantizes upward. Listed twice -- survey puts it under training.
Drynx2020-03inferenceRegression over 600k records: <22s proving, 56kB proof.
Verifiable evaluations of machine learning models using zkSNARKstesting
Efficient Public Verification of Private ML via Regularizationinference
zk-OPMLinferenceHybrid optimistic verification + ZKP for isolated ONNX operators.
awesome-zkml (link collection)infrastructure
Zero-Knowledge AI Inference with High PrecisioninferenceCited by DeepProve [RWB+25]. HIGHEST-PRIORITY follow-up: recent, peer-reviewed, and its 'high precision' angle is the exact accuracy/floating-point axis DeepProve and zkLLM contest. Unknown whether it targets LLMs or general ML -- needs to be pulled and read.
An Efficient and Extensible Zero-knowledge Proof Framework for Neural NetworksinferenceCited by DeepProve. Same core authors as zkGPT (Wenjie Qu, Jiaheng Zhang) -- likely zkGPT's general-NN sibling or precursor. Check whether it is distinct from zkGPT before adding as a separate system.