P089 Zero-knowledge proofs of training and inference
What is the largest model for which a zero-knowledge proof of training or inference is practical, and at what prover overhead?
Zero-knowledge proofs of training and inference.
Terms: zero-knowledge proof: a cryptographic proof that a statement is true (e.g. that a computation was done correctly) while revealing nothing else about the inputs.
Status
Started Inference only: a built, open-source ZK proof system measured on LLaMa-2-13B and OPT-13B (one A100 plus 12 CPU cores): 803 s to prove a 2048-token inference, proof under 200 kB, verification in seconds; impractical at production request rates, and later systems (ZKTorch, zkGPT) remain at GPT-2 to Llama-2-7B scale with minutes per token. Training: no ZK proof of full-LLM backpropagation in native floating point has been shown (Peigné-Lefebvre et al. 2026); their 36-month, single-digit-percent-overhead estimate is a paper estimate, and zkLLM's authors write that ZK proofs of LLM training may pose insurmountable challenges.
"for LLMs boasting 13 billion parameters, our approach enables the generation of a correctness proof for the entire inference process in under 15 minutes."
Posed by
- Reuel, Bucknall et al. (2025), Open Problems in Technical AI Governance, compute sections:
addressing speed constraints will be necessary if such methods are to be applied to larger models
(§5.4.1, p.31, L1691 (Q62/Q63)) (source states the problem but not the quantity the question asks for)
Also raised in: Harack et al. (2025), Verification for International AI Governance (Oxford AIGI); Baker, Kulp, Marks, Brundage & Heim (2025), Verifying International Agreements on AI: Six Layers of Verification for Rules on Large-Scale AI Development and Deployment (RAND WR-A4077-1); Shavit (2023), What Does It Take to Catch a Chinchilla? Verifying Rules on Large-Scale Neural Network Training via Compute Monitoring; Cankaya (2026), A System Overview for Near-Term, Low-Trust AI Compute Verification (v0.2); Cankaya (2026), Bit-Exact AI Inference Verification Without Performance Tradeoffs, arXiv 2606.00279v2; Ansari (2026), Hardware-Level Governance of AI Compute: A Feasibility Taxonomy for Regulatory Compliance and Treaty Verification; Hausenloy & Li (2026), The Compute Verification Post; Rinberg et al. (2025), Verifying LLM Inference to Detect Model Weight Exfiltration; Jin et al. (2026), Trusting What You Cannot See: Auditable Fine-Tuning and Inference for Proprietary AI, arXiv 2603.07466; Peigné-Lefebvre, Nguyen & Wang (2026), Zero Knowledge Verification for Frontier AI Training Is Possible, arXiv 2606.05433; AI 2040 (2026), Get Involved in Verification; AI Security Forum, Requests for Discussion; Dean, Verification Plan (AI 2040); Proofworks, Living Verification Documents.
Sub-questions
| ID | Sub-question | What the sources report |
|---|---|---|
| P089.1 | What is the largest model for which a ZK proof of a full training step (incl. backprop in native float) exists, and at what prover overhead (against a target to be agreed)? | partly answered |
| P089.2 | What is the best ZK inference proving time per token for models across a stated size band (and GPU-seconds per 13B forward pass)? | open |
| P089.3 | Is there any inference-verification scheme for >=30B models with verifier cost <= plain recompute, other than attestation? | a source suggests an answer (not checked) |
| P089.4 | What prover-time multiplier does bit-exact floating-point reduction emulation add inside ZK, and has a ZK proof matched GPU BF16 outputs bit-exactly? | partly answered |
| P089.5 | How much cheaper per MAC is proving INT8 than BF16, and can integer-friendly kernels close the gap? | open |
| P089.6 | Can the BF16/FP32 MAC precompile reach <=50 constraints/op, and what zkVM throughput (constraints/s/GPU) is achieved? | open |
| P089.7 | Does an approximate Freivalds test keep soundness to a stated bound for BF16 GEMM? | open |
| P089.8 | Has an architecture-private zkSNARK for a transformer been demonstrated? | open |
| P089.9 | Does a ZK inference system accept a computationally trivial ('hollow') model as the committed model, and can this be prevented? | open |
| P089.10 | Do ZK or full-TEE verification of sampled blocks scale to modern LLMs? | partly answered |
Sub-questions not yet stated precisely enough to check
| ID | Sub-question |
|---|---|
| P089.11 | What does verification cost at FP8/FP4? |
| P089.12 | Is a hybrid ZKP + hardware approach practical sooner than pure ZK? |
Possibly relevant work found by automated search
Found by AI agents searching for this problem's sub-questions, mostly by keyword and citation in the OpenAlex scholarly database and on the web; a second AI pass reviewed each one, and we spot-checked some by hand. They seemed reasonably related, as far as one can tell from outside each paper's field. Some come from neighbouring fields, such as arms control and financial audit, and bear on the problem by analogy. Not reviewed against the status above; a paper listed here does not change that status. 928 of 1052 quotes checked word for word against the paper's text.
For P089.1: What is the largest model for which a ZK proof of a full training step (incl. backprop in native float) exists, and at what prover overhead (against a target to be agreed)?
For P089.2: What is the best ZK inference proving time per token for models across a stated size band (and GPU-seconds per 13B forward pass)?
For P089.3: Is there any inference-verification scheme for >=30B models with verifier cost <= plain recompute, other than attestation?
For P089.4: What prover-time multiplier does bit-exact floating-point reduction emulation add inside ZK, and has a ZK proof matched GPU BF16 outputs bit-exactly?
For P089.7: Does an approximate Freivalds test keep soundness to a stated bound for BF16 GEMM?
For P089.8: Has an architecture-private zkSNARK for a transformer been demonstrated?
* Full text not reviewed; more likely to contain errors.
Data
- problems.csv — one row per problem
- problem_status.csv — status, scope and evidence per problem
- problem_sources.csv — the posed-by quote and other documents per problem
- leaves.csv — all sub-questions, including those without a checkable form
- leaf_sources.csv — every source row behind each sub-question