Verification of model architecture and training procedures

Show practical methods for verifying properties such as model architecture and training procedures to support formal compliance claims by developers and deployers.

Background

Beyond performance metrics, regulators may require attestation of architectural and training details. Formal verification at model scale remains largely untested and difficult.

Developing methods to verify architecture and training processes would strengthen compliance and trust in declared system properties.

References

Furthermore, verifying properties such as a system's architecture or training procedure remain open questions.

— Open Problems in Technical AI Governance  (2407.14981 - Reuel et al., 2024) in Section 5.3.1 Verification of Model Properties

PA-FL Lite attests that a committed batch satisfies physical invariants over sampled rows, leaving open whether the submitted update was computed strictly from that batch.

— Physics-Attested Federated Learning: Securing Collaborative Anomaly Detection in Critical Water Infrastructure  (2609.34804 - Nijsse et al., 28 Sep 2026) in Section 6, Limitations

Taken together, prior work either targets per-query inference, proves the training trajectory through an expensive backpropagation circuit, or scales only to demonstration-level learners. Efficient checkpoint-level, commit-then-challenge auditing for LLM training outcomes therefore remains largely open.

— zkLLMPoT: Efficient Zero Knowledge Proof of Training for Large Language Models  (2610.08258 - Liang et al., 6 Oct 2026) in Section 1, Introduction