arXiv cs.AI· Junkai Liang, Zhanpeng Guo, Pengfei Wu, Qingni Shen, Jiaheng Zhang, Zhonghai Wu, Haiyang Xue, Shengfang Zhai·· 10 天前AI 评分27
zkLLMPoT：面向大语言模型训练的高效零知识证明框架
zkLLMPoT: Efficient Zero Knowledge Proof of Training for Large Language Models
AI 导读
zkLLMPoT 通过前向评估而非验证优化轨迹，为 LLM 训练检查点提供零知识证明，认证成本与训练迭代次数无关，且无需泄露模型权重或私有训练数据。该框架分两阶段：训练方先固定架构并提交权重，审计方再选取挑战序列，训练方随后证明所提交模型在这些序列上的目标值。在四类模型家族上，1.1-1.5B 参数模型的证明时间为 41-59 秒，13B 模型为 131 秒，序列长度 512 时验证耗时低于半秒。
正文 · 原文
Abstract:Auditing the claimed outcomes of large language model (LLM) training is challenging when model weights and training data are private, while cryptographically proving the full training process is prohibitively expensive at Transformer scale. We present zkLLMPoT, a zero-knowledge framework that certifies auditor-defined properties of a trained checkpoint through forward evaluation rather than verification of its optimization trajectory. zkLLMPoT includes 2 phases: 1) The trainer fixes the architecture and the model weights are committed. Then the auditor selects challenge sequences, preventing the trainer from modifying the checkpoint in response to the audit data. 2) Then the trainer proves the objective value attained by the committed model on those sequences. This formulation makes the certification cost independent of the number of training iterations, without revealing model weights or requiring access to private training data. We build on sumcheck- and lookup-based arguments to certify Transformer computations, while supporting next-token loss and task-specific audit objectives. Across four model families, operator-level benchmarks yield proving times of 41-59 seconds for 1.1-1.5B-parameter models and 131 seconds at 13B for the covered operators, with verification below half a second at a sequence length of 512.
| Comments: | Submitted to ICLR 2027 |
| Subjects: | Artificial Intelligence (cs.AI) |
| MSC classes: | 94A60, 68T50 |
| ACM classes: | E.3; I.2.6; I.2.7 |
| Cite as: | arXiv:2610.08258 [cs.AI] |
| (or arXiv:2610.08258v1 [cs.AI] for this version) | |
| https://doi.org/10.48550/arXiv.2610.08258 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Junkai Liang [view email]
[v1]
Tue, 6 Oct 2026 12:35:49 UTC (112 KB)
来源：arXiv cs.AI · arxiv.org