Pith.

Hardware Architecture · arXiv:2611.17042 v1 · submitted 12 Nov 2026 · 71 references

Infrastructure for Verifiable AI: Where Proof Costs Go and How Hardware Can Help

Yihan Xia, Wenxin Xu, Taotao Wang

The paper says

“Users of artificial intelligence services, including those powered by large language models and other transformer-based models, need assurance that inference and training follow the declared computation without exposing private models or data.”

more of the abstract

Zero-knowledge proofs can provide this assurance, but proof generation remains expensive. Hardware acceleration research often reports gains for individual cryptographic operations, leaving their effect on the full cost of proving neural-network workloads, particularly large transformer models, unclear. We introduce a four-layer framework linking verifiable AI workloads, proof systems, cryptographic kernels and hardware platforms, use it to review the literature, and add original measurements of six provers and seven kernels to identify which costs matter in practice. Our measurements show that the dominant cost changes with the proof system and input length. In low-rank fine-tuning, the cost of opening commitments to fixed model weights remains nearly constant across input lengths from 64 to 2048, while the work associated with intermediate values grows. A 13-billion-parameter fine-tuning configuration exceeds the capacity of a 40-gigabyte graphics processor but completes on an 80-gigabyte device, showing why capacity must be assessed alongside execution time. Our kernel profiling further identifies instruction issue and synchronization limits that a bandwidth-only model misses. We use these findings to assess graphics processors, reconfigurable logic, custom chips and existing artificial intelligence accelerators. The resulting research agenda connects proof-friendly numeric formats, reuse of fixed-weight commitments, heterogeneous execution and shared benchmarks to the measured bottlenecks. Effective acceleration requires selecting hardware and proof-system optimizations around the cost profile of the target workload.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 13 Nov 2026 · what was read was not recorded · not checked by a person

It can be wrong. Challenge this review or reply to it.

Careful measurement study of ZKP costs for AI workloads; useful framework, honest about its sumcheck-only scope. This line is wrong

the referee pass did not run

How often objections from reviews like this one are right: not yet measured

Are you an author? What you can do.

Not checked: figures, code, what the 71 references say.

What would settle it

If one wanted to test the generalizability of the cost trends beyond sumcheck-based provers, one could implement a hash-based prover (e.g., using Brakedown or Basefold commitments) for a small transformer block (e.g., 1 layer, 128 hidden dimension) and measure the cost breakdown using the paper's taxonomy. If the dominant cost categories differ substantially from those reported for sumcheck-based provers (e.g., if encoding and hashing dominate rather than sumcheck and lookup), then the paper's claim that 'no single cost dominates' would need qualification for that family. However, the paper already acknowledges this as a limitation and does not claim the trends hold for hash-based systems.

from the skeptic's reading

How the review went

3 readings by deepseek-v4-flash · 13 Nov 2026

  1. Reader holds up confidence high
  2. Skeptic agrees 9 s · 29,183 tokens read, 687 written
  3. Editor wrote the verdict above 8 s · 29,939 tokens read, 498 written

Not run on this paper: the referee, the simulated authors' defence.

Pith did not check this review's objections against the paper.

The review in full

The review in brief, machine-written

This paper is worth reading if you work on hardware for zero-knowledge proofs, especially for AI workloads. It does two things well. First, it provides a clean four-layer framework (workload → proof system → kernel → hardware) that organizes a messy literature. Second, it contributes original, careful measurements of six provers and seven kernels on real A100 GPUs, with a common cost taxonomy and roofline analysis using measured roofs. The key findings are that no single kernel dominates across systems, that fixed-weight opening costs are constant with input length, and that memory capacity (not just bandwidth) is a first-order constraint. The kernel profiling also corrects some folklore: low-degree sumcheck is instruction-issue-bound, not memory-bound, on this hardware.

The soft spots are real but contained. The measurements are mostly restricted to sumcheck-based provers; the authors say this explicitly. The non-sumcheck point is a CPU-only MLP block, and the commitment comparison is on one transformer block. So the framework's predictions for hash-based or lattice-based systems at scale remain untested. The engineering waterfall (Section IV-D) is interesting but inconclusive—the authors themselves say it does not bound the gains from hardware. The paper also leans heavily on the reader accepting that the four-layer framework is a contribution rather than a taxonomy; I think it is a useful organizational tool, but it is not a theoretical result.

Who is this for? Hardware architects working on ZKP accelerators, and system builders who need to know where to optimize. The paper does not overclaim; it presents its findings as a starting point for workload-aware design. The measurements are reproducible (public code, documented methodology).

Recommendation: This deserves a serious peer review. The measurements are solid, the limitations are stated, and the framework is useful even if incomplete. I would accept it with minor revisions—mostly asking for a clearer separation of what is measured versus what is extrapolated, and a note on whether the roofline analysis would change on H100 or AMD hardware. I would not desk-reject it.

The strongest objection

The paper is a measurement study and framework paper, not a derivation of a new cryptographic scheme or a claim of a universal law. Its central claim is that the cost profile of verifiable AI proof generation shifts with the proof system and input length, that fixed-weight opening costs are stable, and that memory capacity is a first-order constraint. The authors explicitly acknowledge the main limitation: full-scale coverage is mostly restricted to sumcheck-based provers, and the non-sumcheck point is a CPU-only MLP block. They do not overclaim generalizability beyond the measured systems. The measurements are carefully conducted, the limitations are clearly stated, and the conclusions are appropriately hedged. The paper's own sentences confirm that the authors are aware of the scope and do not assert universal applicability. Therefore, no load-bearing concern about the central argument arises from the evidence presented. The weakest assumption identified by the reader (generalizability beyond sumcheck-based provers) is the same limitation the paper itself states, and it is not a flaw in the argument but a scope condition. The paper's contribution is a well-executed measurement study and a useful framework, and it does not make unsupported claims.

The reader's verdict is ACCEPT with HIGH confidence, and the weakest assumption identified is the generalizability beyond sumcheck-based provers. The paper itself states this limitation explicitly in Section IV-A: 'Full-scale coverage is mostly restricted to sumcheck-based provers; the non-sumcheck point is a CPU-only multilayer-perceptron block, and the commitment comparison is made on one transformer block.' The paper does not claim that its cost breakdowns apply to all proof systems; it presents them as measurements of specific systems and draws conclusions that are appropriately scoped. The central claim—that the dominant cost shifts with proof system and input length, that fixed-weight opening costs are stable, and that memory capacity is a first-order constraint—is well-supported by the measurements presented. The kernel profiling and roofline analysis are careful and provide new insights. The hardware review is thorough and identifies gaps in the literature. The paper is a significant contribution to understanding where costs go in verifiable AI proof generation, and it does not overclaim. Therefore, no adjustment to the verdict is needed. The agreement with the reader is 'agree' because the reader's weakest assumption is the same limitation the paper acknowledges, and the reader's verdict is appropriate.

Pith's reading of the paper (a machine's summary, not the authors')

Users of AI services need to trust that the computation was done correctly, without exposing private models or data. Zero-knowledge proofs can provide this cryptographic guarantee, but generating the proof is very expensive. This paper studies where that cost actually goes when proving large language models and other neural networks. The authors introduce a four-layer framework that connects the AI workload (like inference or fine-tuning), the proof system (like GKR or Plonk), the cryptographic kernels (like multi-scalar multiplication, sumcheck, or hashing), and the hardware platform (like GPUs, FPGAs, or ASICs). They then measure six different provers and seven kernels on real hardware. The key findings are: (1) no single cost dominates; the largest share changes depending on the proof system and the input length. (2) The cost of opening commitments to fixed model weights stays nearly constant as the input length grows, while the work for intermediate values grows. (3) Memory capacity, not just bandwidth, is a first-order limit: a 13-billion-parameter fine-tuning configuration exceeds a 40 GB GPU but fits on 80 GB. (4) Kernel profiling shows that some kernels thought to be bandwidth-bound are actually limited by instruction issue or synchronization on GPUs. The paper uses these findings to assess GPUs, FPGAs, ASICs, and AI accelerators, and proposes a research agenda including proof-friendly numeric formats, reuse of fixed-weight commitments, heterogeneous execution, and shared benchmarks.

Core claim. The central claim is that the cost of generating zero-knowledge proofs for AI workloads is not dominated by a single kernel but shifts with the proof system and input length, and that memory capacity (not just bandwidth) is a first-order constraint. The paper states: 'Our measurements show that (i) no single cost dominates, as the largest share shifts with the proof system and the input length; (ii) the cost of opening fixed weights does not grow with the input length; and (iii) memory capacity limits what can be proved before data movement limits how fast.' (Section I, paragraph 4). If correct, this means hardware acceleration must be workload-aware, targeting the specific cost profile of the target proof system and scale.

The premise it rests on. The measurements are limited to sumcheck-based provers for large models (zkLLM, VeriLoRA, DeepProve) and a few smaller points (Expander block, ZKTorch, VeriDP). The paper acknowledges this: 'Full-scale coverage is mostly restricted to sumcheck-based provers; the non-sumcheck point is a CPU-only multilayer-perceptron block, and the commitment comparison is made on one transformer block.' (Section IV-A, Limitations). The claim that 'no single cost dominates' and the specific cost shares are therefore only directly supported for the sumcheck family. The paper's conclusions about hardware acceleration priorities may not generalize to hash-based or lattice-based proof systems at scale, which are not measured end-to-end. The authors note that no hash-based prover runs one end to end (Section IV-A). This is a structural limitation of the current literature, not a flaw in the paper, but it means the framework's predictions for those systems remain untested.

Check this review

Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.

How often reviews like this one are right

Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.

How often objections from reviews like this one are right: not yet measured

Every value above, in the review's public record (JSON)

Reviews by people

This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.

Request a human review

How Pith checks papers · The full audit of this paper

The paper

Abstract

Users of artificial intelligence services, including those powered by large language models and other transformer-based models, need assurance that inference and training follow the declared computation without exposing private models or data. Zero-knowledge proofs can provide this assurance, but proof generation remains expensive. Hardware acceleration research often reports gains for individual cryptographic operations, leaving their effect on the full cost of proving neural-network workloads, particularly large transformer models, unclear. We introduce a four-layer framework linking verifiable AI workloads, proof systems, cryptographic kernels and hardware platforms, use it to review the literature, and add original measurements of six provers and seven kernels to identify which costs matter in practice. Our measurements show that the dominant cost changes with the proof system and input length. In low-rank fine-tuning, the cost of opening commitments to fixed model weights remains nearly constant across input lengths from 64 to 2048, while the work associated with intermediate values grows. A 13-billion-parameter fine-tuning configuration exceeds the capacity of a 40-gigabyte graphics processor but completes on an 80-gigabyte device, showing why capacity must be assessed alongside execution time. Our kernel profiling further identifies instruction issue and synchronization limits that a bandwidth-only model misses. We use these findings to assess graphics processors, reconfigurable logic, custom chips and existing artificial intelligence accelerators. The resulting research agenda connects proof-friendly numeric formats, reuse of fixed-weight commitments, heterogeneous execution and shared benchmarks to the measured bottlenecks. Effective acceleration requires selecting hardware and proof-system optimizations around the cost profile of the target workload.

The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.

abstract compiled · 2026-11-13

The authors' main claim, as the classifier picked it out 0.98

“Effective acceleration requires selecting hardware and proof-system optimizations around the cost profile of the target workload.”

How the main claim is stated: asserted 1.00

What the abstract itself states, as probabilities: claims first 0.07 · states limitation 0.34 · human participants 0.02 · controlled comparison 0.10 · randomized assignment 0.01 · states quantitative result 0.96 · compares with prior results 0.67 · mentions code or data availability 0.03

The abstract, sentence by sentence

This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.

SentenceRoleThe authors' own workStates the main conclusionStrength
Users of artificial intelligence services, including those powered by large language models and other transformer-based models, need assurance that inference and training follow the declared computation without exposing private models or data. background 1.00 0.04 0.02 asserted 0.98
Zero-knowledge proofs can provide this assurance, but proof generation remains expensive. background 0.92 0.05 0.03 asserted 0.88
Hardware acceleration research often reports gains for individual cryptographic operations, leaving their effect on the full cost of proving neural-network workloads, particularly large transformer models, unclear. gap 1.00 0.05 0.02 asserted 0.95
We introduce a four-layer framework linking verifiable AI workloads, proof systems, cryptographic kernels and hardware platforms, use it to review the literature, and add original measurements of six provers and seven kernels to identify which costs matter in practice. contribution 0.94 0.96 0.07 asserted 0.99
Our measurements show that the dominant cost changes with the proof system and input length. result 1.00 0.96 0.31 asserted 1.00
In low-rank fine-tuning, the cost of opening commitments to fixed model weights remains nearly constant across input lengths from 64 to 2048, while the work associated with intermediate values grows. result 1.00 0.93 0.08 asserted 1.00
A 13-billion-parameter fine-tuning configuration exceeds the capacity of a 40-gigabyte graphics processor but completes on an 80-gigabyte device, showing why capacity must be assessed alongside execution time. result 0.94 0.91 0.06 asserted 1.00
Our kernel profiling further identifies instruction issue and synchronization limits that a bandwidth-only model misses. result 0.90 0.94 0.04 asserted 1.00
We use these findings to assess graphics processors, reconfigurable logic, custom chips and existing artificial intelligence accelerators. contribution 0.68 0.83 0.05 asserted 0.91
The resulting research agenda connects proof-friendly numeric formats, reuse of fixed-weight commitments, heterogeneous execution and shared benchmarks to the measured bottlenecks. future work 0.94 0.81 0.14 asserted 0.98
Effective acceleration requires selecting hardware and proof-system optimizations around the cost profile of the target workload. interpretation 0.94 0.52 0.75 asserted 1.00

What kind of paper the classifier reports this is: empirical study 0.98 · evidence from observation or experiment 1.00

What was examined, and what was not

Sources read

  • arXiv listing abstract v1

Not examined

  • full text
  • figures and tables
  • references and citing sentences
  • relations to other papers

Levels run

  • P abstract

Not run

  • S sections
  • T body sentences
  • C citations
  • R cross-paper relations

How the labels were made

Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-11-13: 1 request, 8,286 input tokens.

Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON

Figures (8)
Figure 1 of the paper
Figure 1. Verifiable AI as a service. The provider attaches a zero-knowledge
Figure 2 of the paper
Figure 2. The three building blocks that dominate proofs of neural networks.
Figure 3 of the paper
Figure 3. Four-layer framework. Each verifiable AI workload (Layer 1) is expressed in a proof system (Layer 2), whose cost is concentrated in a few
Figure 4 of the paper
Figure 4. Throughput of the seven Layer-3 kernels (log scale) at
Figure 5 of the paper
Figure 5. Roofline of the kernels at 2 24 elements on the A100: thread-level integer-instruction rate, summed over the ALU and FMA pipelines, against integer instructions per DRAM byte. The diagonal is the measured copy bandwidth (≈1.36 TB/s); the solid and dashed lines are the measured single￾pipe (≈9.66×1012 inst/s) and dual-pipe (≈1.91×1013) issue limits, so points may lie above the solid line. Left of the ridge point (≈7 inst/byte) bandwidth can limit a kernel, right of it instruction issue; a point far below both is limited by dependencies, synchronization or occupancy. Labels: sc…
Figure 6 of the paper
Figure 6. Share of proof time by cost category for each measured (workload, prover, scale) cell; bar tops give the median proof time and the lower strip shows
Figure 7 of the paper
Figure 7. Engineering-optimization waterfall for zkLLM on W1 (7B, one layer,
Figure 8 of the paper
Figure 8. Roadmap toward practical verifiable AI. Numbers refer to the open

Figures are the authors', reproduced from the paper.

References (71)
  1. [1] Y . Sun, Y . Li, Y . Zhang, Y . Jin, and H. Zhang, “SVIP: Towards Verifiable Inference of Open-source Large Language Models,” arXiv:2410.22307, 2024 on PitharXiv
  2. [2] J. M. Ong, M. D. Ferrante, A. Pazdera, R. Garner, S. Jaghouar, M. Basra, M. Ryabinin, and J. Hagemann, “TOPLOC: A Locality Sensitive Hash- ing Scheme for Trustless Verifiable Inference,” arXiv:2501.16007, 2025 on PithPith's reviewarXiv
  3. [3] H. Jin, C. Zhang, H. Yu, S. Shi, N. Zhang, Y . T. Hou, and W. Lou, “Trusting What You Cannot See: Auditable Fine-Tuning and Inference for Proprietary AI,” arXiv:2603.07466, 2026 on PithPith's reviewarXiv
  4. [4] H. Akgul, D. Borg, A. Berisha, A. Rahimova, A. Novak, and M. Petrov, “Verifiable Fine-Tuning for LLMs: Zero-Knowledge Training Proofs Bound to Data Provenance and Policy,” arXiv:2510.16830, 2025 on PitharXiv
  5. [5] B. Abdolmaleki, A. R. Asadi, V . R. Asadi, S. Köpsell, B. Mohee, N. Roustaeifar, and M. Zarezadeh, “VeriDP: Verifiable Differentially Private Training,”Proceedings on Privacy Enhancing Technologies, vol. 2026, no. 3, pp. 48–64, 2026 on Pith
  6. [6] H. Jia, M. Yaghini, C. A. Choquette-Choo, N. Dullerud, A. Thudi, V . Chandrasekaran, and N. Papernot, “Proof-of-Learning: Definitions and Practice,” inProc. IEEE Symposium on Security and Privacy (SP), 2021, pp. 1039–1056 on Pith
  7. [7] Z. Chen, M. Wu, M. Heinrich, J. Zeng, H. Lan, T. Zhang, and R. Tan, “Communication-Efficient Verifiable Attention for LLM Inference,” arXiv:2606.16352, 2026 on PitharXiv
  8. [8] A. Arun, A. S. Arnaud, A. Titov, B. Wilcox, V . Kolobaric, M. Brinkmann, O. Ersoy, B. Fielding, and J. Bonneau, “Verde: Ver- ification via Refereed Delegation for Machine Learning Programs,” arXiv:2502.19405, 2025 on PithPith's reviewarXiv
  9. [9] H. Sun, J. Li, and H. Zhang, “zkLLM: Zero Knowledge Proofs for Large Language Models,” inProc. ACM Conference on Computer and Communications Security (CCS), 2024, pp. 4405–4419 on Pith
  10. [10] W. Qu, Y . Sun, X. Liu, T. Lu, Y . Guo, K. Chen, and J. Zhang, “zkGPT: An Efficient Non-interactive Zero-knowledge Proof Framework for LLM Inference,” inProc. 34th USENIX Security Symposium, 2025, pp. 2045– 2063 on Pith
  11. [11] N. Gailly, I. Hishon-Rezaizadeh, T. Liu, N. Mainardi, D. Papadopoulos, C. Papamanthou, C. Pappas, S. Srinivasan, Z. Youell, and Y . Zhang, “DeepProve: Verifiable End-to-End Large Language Model Inference,” inProc. ACM Conference on Computer and Communications Security (CCS), 2... on Pith
  12. [12] L. Wang, H. Lou, C. Li, Y . Yu, and Y . Hu, “zkAgent: Verifiable LLM Agent Execution via One-Shot Transcript Proofs,” inProc. ACM Conference on Computer and Communications Security (CCS), 2026, To appear on Pith
  13. [13] G. Liao, T. Wang, S. Zhang, J. Zhang, L. Shi, and D. Tao, “VeriLoRA: Fine-Tuning Large Language Models with Verifiable Security via Zero- Knowledge Proofs,” inProc. Network and Distributed System Security Symposium (NDSS), 2026 on Pith
  14. [14] Q. Hu, C. Peng, Y . Dai, M. Luo, and D. He, “A Survey on the Implementation of zk-SNARK Hardware Accelerators,”Chinese Journal of Electronics, vol. 35, no. 3, pp. 1180–1194, 2026 on Pith
  15. [15] Y . Zhang, S. Wang, X. Zhang, J. Dong, X. Mao, F. Long, C. Wang, D. Zhou, M. Gao, and G. Sun, “PipeZK: Accelerating Zero-Knowledge Proof with a Pipelined Architecture,” inProc. ACM/IEEE International Symposium on Computer Architecture (ISCA), 2021, pp. 416–428 on Pith
  16. [16] W. Ma, Q. Xiong, X. Shi, X. Ma, H. Jin, H. Kuang, M. Gao, Y . Zhang, H. Shen, and W. Hu, “GZKP: A GPU Accelerated Zero-Knowledge Proof System,” inProc. ACM International Conference on Architectural Support for Programming Languages and Operating Systems (ASPLOS), 2023, pp. 340–353 on Pith
  17. [17] T. Lu, C. Wei, R. Yu, C. Chen, W. Fang, L. Wang, Z. Wang, and W. Chen, “cuZK: Accelerating Zero-Knowledge Proof with A Faster Parallel Multi-Scalar Multiplication Algorithm on GPUs,”IACR Trans- actions on Cryptographic Hardware and Embedded Systems, pp. 194– 220, 2023 on Pith
  18. [18] X. Pottier, T. D. Ruijter, J. Bertels, W. Legiest, M. V . Beirendonck, and I. Verbauwhede, “OPTIMSM: FPGA hardware accelerator for Zero- Knowledge MSM,”IACR Transactions on Cryptographic Hardware and Embedded Systems, vol. 2025, no. 2, pp. 489–510, 2025 on Pith
  19. [19] A. Daftardar, B. Reagen, and S. Garg, “SZKP: A Scalable Accelerator Architecture for Zero-Knowledge Proofs,” inProc. International Con- ference on Parallel Architectures and Compilation Techniques (PACT), 2024, pp. 271–283 on Pith
  20. [20] T. Verma, Y . Yuan, N. Talati, and T. Austin, “ZKProphet: Understanding Performance of Zero-Knowledge Proofs on GPUs,” inProc. IEEE International Symposium on Workload Characterization (IISWC), 2025, pp. 142–154 on Pith
show 51 more
  1. [21] K. Farheen and N. G. Tsoutsos, “Chimera: A Hybrid GPU Backend for Sumcheck Acceleration in Zero-Knowledge Provers,” Cryptology ePrint Archive, Paper 2026/1388, 2026 on Pith
  2. [22] T. Lu, Y . Chen, Z. Wang, X. Wang, W. Chen, and J. Zhang, “BatchZK: A Fully Pipelined GPU-Accelerated System for Batch Generation of Zero-Knowledge Proofs,” inProc. ACM International Conference on Architectural Support for Programming Languages and Operating Sys- tems (ASPLOS)... on Pith
  3. [23] A. Putra, C. M. Duong, A. Q. Pham, and J.-Y . Kim, “ZK-Flex: A Flexible and Scalable Framework for Accelerating Zero-Knowledge Proofs,” inProc. ACM/IEEE Design Automation Conference (DAC), 2026, pp. 1–7 on Pith
  4. [24] Z. Yang, L. Zhao, P. Li, H. Liu, K. Li, B. Zhao, D. Meng, and R. Hou, “LegoZK: A Dynamically Reconfigurable Accelerator for Zero- Knowledge Proof,” inProc. IEEE International Symposium on High- Performance Computer Architecture (HPCA), 2025, pp. 113–126 on Pith
  5. [25] R. Lavin, X. Liu, H. Mohanty, L. Norman, G. Zaarour, and B. Krish- namachari, “A Survey on the Applications of Zero-Knowledge Proofs,” arXiv:2408.00243, 2024 on PithPith's reviewarXiv
  6. [26] Z. Peng, C. Zhao, T. Wang, G. Liao, Z. Lin, Y . Liu, B. Cao, L. Shi, Q. Yang, and S. Zhang, “A survey of zero-knowledge proof based verifiable machine learning,”Artificial Intelligence Review, vol. 59, no. 7, 2026 on Pith
  7. [27] T. Wang, Y . Jin, Q. Yang, Y . Xia, L. Shi, and S. Zhang, “Zero-Knowledge Federated Learning: A New Trustworthy and Privacy-Preserving Dis- tributed Learning Paradigm,”IEEE Communications Magazine, vol. 64, no. 5, pp. 172–178, 2026 on Pith
  8. [28] A. Y . L. Kei and S. S. M. Chow, “SoK: Private Transformer Inference Across Systems, Models, and Cryptography,” Cryptology ePrint Archive, Paper 2026/2005, 2026 on Pith
  9. [29] A. A. Badawi, A. Alexandru, Y . Polyakov, and V . Vaikuntanathan, “SoK: Private LLM Inference using Approximate Homomorphic Encryption,” Cryptology ePrint Archive, Paper 2026/935, 2026 on Pith
  10. [30] C. Skatharoudis, “SoK: Hash-Based Polynomial Commitments and Low- Degree Tests: From FRI to Basefold, STIR, and WHIR,” Cryptology ePrint Archive, Paper 2026/1367, 2026 on Pith
  11. [31] J. Liang, Z. Guo, P. Wu, Q. Shen, J. Zhang, Z. Wu, H. Xue, and S. Zhai, “zkLLMPoT: Efficient Zero Knowledge Proof of Training for Large Language Models,” arXiv:2610.08258, 2026 on PithPith's reviewarXiv
  12. [32] Attestable, “Proving LLMs at scale,” https://attestable.com/blog/ proving-llms-scale, 2026 on Pith
  13. [33] E. Ben-Sasson, I. Bentov, Y . Horesh, and M. Riabzev, “Fast Reed- Solomon interactive oracle proofs of proximity,” inProc. 45th Interna- tional Colloquium on Automata, Languages, and Programming (ICALP), ser. LIPIcs, vol. 107, 2018, pp. 14:1–14:17 on Pith
  14. [34] A. Golovnev, J. Lee, S. Setty, J. Thaler, and R. S. Wahby, “Brakedown: Linear-Time and Field-Agnostic SNARKs for R1CS,” inAdvances in Cryptology (CRYPTO), 2023, pp. 193–226 on Pith
  15. [35] H. Zeilberger, B. Chen, and B. Fisch, “BaseFold: Efficient Field- Agnostic Polynomial Commitment Schemes from Foldable Codes,” in Advances in Cryptology (CRYPTO), 2024, pp. 138–169 on Pith
  16. [36] A. Kate, G. M. Zaverucha, and I. Goldberg, “Constant-Size Com- mitments to Polynomials and Their Applications,” inAdvances in Cryptology (ASIACRYPT), 2010, pp. 177–194 on Pith
  17. [37] R. S. Wahby, I. Tzialla, A. Shelat, J. Thaler, and M. Walfish, “Doubly- Efficient zkSNARKs Without Trusted Setup,” inProc. IEEE Symposium on Security and Privacy (SP), 2018, pp. 926–943 on Pith
  18. [38] C. Lund, L. Fortnow, H. Karloff, and N. Nisan, “Algebraic methods for interactive proof systems,”Journal of the ACM, vol. 39, no. 4, pp. 859–868, 1992 on Pith
  19. [39] S. Goldwasser, Y . T. Kalai, and G. N. Rothblum, “Delegating Compu- tation: Interactive Proofs for Muggles,”Journal of the ACM, vol. 62, no. 4, pp. 1–64, 2015 on Pith
  20. [40] U. Haböck, “Multivariate lookups based on logarithmic derivatives,” Cryptology ePrint Archive, Paper 2022/1530, 2022. INFRASTRUCTURE FOR VERIFIABLE AI 16 on Pith
  21. [41] A. Gabizon, Z. J. Williamson, and O. Ciobotaru, “PLONK: Permuta- tions over Lagrange-bases for Oecumenical Noninteractive arguments of Knowledge,” Cryptology ePrint Archive, Paper 2019/953, 2019 on Pith
  22. [42] L. Grassi, D. Khovratovich, and M. Schofnegger, “Poseidon2: A Faster Version of the Poseidon Hash Function,” inProgress in Cryptology (AFRICACRYPT), 2023, pp. 177–203 on Pith
  23. [43] J. O’Connor, J.-P. Aumasson, S. Neves, and Z. Wilcox- O’Hearn, “BLAKE3: One function, fast everywhere,” https://github.com/BLAKE3-team/BLAKE3-specs, 2020 on Pith
  24. [44] B. E. Diamond and J. Posen, “Succinct Arguments over Towers of Bi- nary Fields,” inAdvances in Cryptology (EUROCRYPT), 2025, pp. 93– 122 on Pith
  25. [45] T. Lu, J. Zhang, Y . Guo, X. Liu, W. Qu, Z. Wang, W. Chen, and J. Zhang, “TensorZKP: Repurposing GPU tensor cores for High-Performance zero-knowledge proofs,” inProc. 35th USENIX Security Symposium, 2026, pp. 5849–5867 on Pith
  26. [46] F. Hirner, F. Krieger, C. Piber, and S. S. Roy, “Accelerating Hash- Based Polynomial Commitment Schemes with Linear Prover Time,” IACR Transactions on Cryptographic Hardware and Embedded Systems, vol. 2025, no. 4, pp. 341–385, 2025 on Pith
  27. [47] M. Li, Y . Yu, B. Wang, X. Fan, and S. Deng, “ZKPoG: Accelerat- ing WitGen-Incorporated End-to-End Zero-Knowledge Proof on GPU,” Cryptology ePrint Archive, Paper 2025/765, 2025 on Pith
  28. [48] M. Srivastava, S. Arora, and D. Boneh, “Optimistic Verifiable Training by Controlling Hardware Nondeterminism,” inAdvances in Neural Information Processing Systems 37 (NeurIPS), 2024, pp. 95 639–95 661 on Pith
  29. [49] B. Chen, B. Bünz, D. Boneh, and Z. Zhang, “HyperPlonk: Plonk with Linear-Time Prover and High-Degree Custom Gates,” inAdvances in Cryptology (EUROCRYPT), 2023, pp. 499–530 on Pith
  30. [50] W. Beullens and G. Seiler, “LaBRADOR: Compact Proofs for R1CS from Module-SIS,” inAdvances in Cryptology (CRYPTO), 2023, pp. 518–548 on Pith
  31. [51] J. Mo, A. Daftardar, B. GaneshKumar, K. Guo, H. Wang, B. Bunz, S. Garg, and B. Reagen, “When Proofs Meet Hardware: Comparing NTT and SumCheck in Zero-Knowledge Systems,” arXiv:2606.16146, 2026 on PitharXiv
  32. [52] K. Inbasekar, Y . Shekel, and M. Asa, “ICICLE v2: Polynomial API for Coding ZK Provers to Run on Specialized Hardware,” Cryptology ePrint Archive, Paper 2024/973, 2024 on Pith
  33. [53] S. A. Butt, B. Reynolds, V . Ramamurthy, X. Xiao, P. Chu, S. Sharifian, S. Gribok, and B. Pasca, “if-ZKP: Intel FPGA-Based Acceleration of Zero Knowledge Proofs,” inProc. IEEE International Symposium on Field-Programmable Custom Computing Machines (FCCM), 2024, pp. 212–212 on Pith
  34. [54] J. Tong, J. Dang, S. Langowski, T. Huang, A. Ali, J. Kun, J. Jiang, S. Devadas, and T. Krishna, “Enabling AI ASICs for Zero Knowledge Proof,” inProc. ACM/IEEE Design Automation Conference (DAC), 2026, pp. 1–7 on Pith
  35. [55] K. Wang, Y . Yanggong, X. Yang, C. Baixiao, and L. Wang, “PipeSC: A Resource-efficient and Pipelined Hardware Accelerator for Sumcheck Protocol,” Cryptology ePrint Archive, Paper 2026/691, 2026 on Pith
  36. [56] A. Daftardar, J. Mo, J. Ah-kiow, B. Bünz, R. Karri, S. Garg, and B. Reagen, “Need for zkSpeed: Accelerating HyperPlonk for Zero- Knowledge Proofs,” inProc. ACM/IEEE International Symposium on Computer Architecture (ISCA), 2025, pp. 1986–2001 on Pith
  37. [57] A. Daftardar, J. Mo, J. Ah-kiow, B. Bünz, S. Garg, and B. Reagen, “zkPHIRE: A Programmable Accelerator for ZKPs over HIgh-degRee, Expressive Gates,” inProc. IEEE International Symposium on High- Performance Computer Architecture (HPCA), 2026, pp. 1–15 on Pith
  38. [58] J. Mo, A. Daftardar, J. Ah-Kiow, K. Guo, B. Bünz, S. Garg, and B. Reagen, “MTU: The Multifunction Tree Unit for Accelerating Zero- Knowledge Proofs,” inProc. International Workshop on Hardware and Architectural Support for Security and Privacy (HASP), 2025, pp. 19– 27 on Pith
  39. [59] N. Samardzic, S. Langowski, S. Devadas, and D. Sanchez, “Accelerating Zero-Knowledge Proofs Through Hardware-Algorithm Co-Design,” in Proc. IEEE/ACM International Symposium on Microarchitecture (MI- CRO), 2024, pp. 366–379 on Pith
  40. [60] P. Peigné, K. Nguyen, and P. Wang, “Proposals and challenges for zero- knowledge verification of frontier AI training,” arXiv:2606.05433, 2026 on PithPith's reviewarXiv
  41. [61] D. Boneh, J. Drake, B. Fisch, and A. Gabizon, “Efficient polynomial commitment schemes for multiple points and polynomials,” Cryptology ePrint Archive, Paper 2020/081, 2020 on Pith
  42. [62] Z. Wang, P. Ma, Z. Xue, Y . Zhou, Q. Zhang, X. Zhang, and S. Wang, “ZK-Value: A Practical Zero-Knowledge System for Verifiable Data Valuation,” arXiv:2605.03581, 2026 on PithPith's reviewarXiv
  43. [63] Polyhedra Network, “Expander: An open-source GKR prover,” https: //github.com/PolyhedraZK/Expander, 2026 on Pith
  44. [64] B.-J. Chen, L. Tang, and D. Kang, “ZKTorch: Compiling ML In- ference to Zero-Knowledge Proofs via Parallel Proof Accumulation,” arXiv:2507.07031, 2025 on PithPith's reviewarXiv
  45. [65] Supranational, “sppark: Zero-knowledge template library,” https://github. com/supranational/sppark, 2026 on Pith
  46. [66] M. J. Jo, “A Separation Principle for Lookup-Based zkML: Activation- Function Structure Cannot Reduce Per-Lookup Proving Cost,” Cryptol- ogy ePrint Archive, Paper 2026/1390, 2026 on Pith
  47. [67] T. Xie, T. Lu, Z. Fang, S. Wang, Z. Zhang, Y . Jia, D. Song, and J. Zhang, “zkPyTorch: A Hierarchical Optimized Compiler for Zero-Knowledge Machine Learning,” Cryptology ePrint Archive, Paper 2025/535, 2025 on Pith
  48. [68] W. Benno, A. Centelles, A. Douchet, and K. Gibran, “Jolt Atlas: Verifiable Inference via Lookup Arguments in Zero Knowledge,” arXiv:2602.17452, 2026 on PitharXiv
  49. [69] C. Wang and M. Gao, “UniZK: Accelerating Zero-Knowledge Proof with Unified Hardware and Flexible Kernel Mapping,” inProc. ACM International Conference on Architectural Support for Programming Languages and Operating Systems (ASPLOS), 2025, pp. 1101–1117 on Pith
  50. [70] N. Cankaya, “Bit-Exact AI Inference Verification Without Performance Tradeoffs,” arXiv:2606.00279, 2026 on PithPith's reviewarXiv
  51. [71] C. Gong, B. Liu, and M. Li, “Hollow-LLM Attack: Computationally Trivial Weights in Zero-Knowledge Verification of LLM Inference,” in Proc. IEEE Symposium on Security and Privacy (SP), 2026, pp. 4339– 4355 on Pith

Challenges and discussion

Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.

No comments yet.

Sign in to comment