2026-10-23 · ArXiv
arXiv:2606.05433v3 Announce Type: replace
Abstract: Proposals for Frontier AI rules often use cumulative training compute as the main criterion for identifying high-impact models. Recent governance analyses call zero-knowledge proofs promising but currently impractical at frontier scale [26, 4]. We argue this barrier reflects the current paradigm, not a fundamental limit. Our frontier dense-pre-training design combines a pre-committed specification, inter-node network observations, and live Merkle commitments to intermediate computation. Our reference zero-knowledge virtual machine (zkVM) provides a succinct proof layer with native BF16/FP32 precompiles. We qualitatively compare zkVMs, custom circuits, and direct interactive audits by transferability, soundness, and cost. Public randomness beacons are a separate audit-selection choice.
At challenged positions, proofs check the committed computation against declared BF16/FP32 semantics bit-for-bit. A private training specification targets model-architecture confidentiality. Genesis proofs cover initialisation; then, in-training step proofs verify selected steps across the run. Ex-ante attestations enforce policy-relevant claims as running invariants on every proved step. Together they create a governance-auditable training record. Detection-probability formulas quantify sampling guarantees from the sampling budget and the share of computation altered. We give numerical examples and explain what sampling does not cover.
We note, however, that weight smuggling could reuse or merge weights trained elsewhere without charging that work to the current run. We then argue that regimes with global training budgets are more robust than regimes with per-run FLOP caps. This advantage depends on counting all contributing training. We catalogue remaining research and engineering problems for external contributors.
原文链接：https://arxiv.org/abs/2606.05433
查看更多资讯请见 AI 行业资讯。