Mengyuan (Marvin) Li

Research direction

Verifiable AI

Mengyuan Li’s research on verifiable AI builds mechanisms that improve the integrity and auditability of deployed models. It focuses on verifying LLM inference and on how proof-based AI systems fail. The work connects verifiable AI with trustworthy machine learning systems and privacy-preserving oversight.

Topics: AI verification · zero-knowledge verification of LLM inference · trustworthy AI systems · model oversight

Papers

  1. IEEE S&P 2026 Hollow-LLM Attack: Computationally Trivial Weights in Zero-Knowledge Verification of LLM Inference Chen Gong, Beijie Liu, Mengyuan Li 2026 IEEE Symposium on Security and Privacy (S&P 2026)

    Shows that a valid zero-knowledge proof of LLM inference is not proof of large-model execution: with “ghost weights”, a provider keeps the declared architecture and passes verification while paying only small-model cost.

    DOI · PDF · arXiv · BibTeX
  2. ASPLOS 2026 WAVE: Leveraging Architecture Observation for Privacy-Preserving Model Oversight Haoxuan Xu*, Chen Gong*, Beijie Liu*, Haizhong Zheng, Beidi Chen, Mengyuan Li 31st ACM International Conference on Architectural Support for Programming Languages and Operating Systems (ASPLOS 2026)

    Explores privacy-preserving oversight of model execution: how to monitor and reason about model behavior in a trustworthy way.

    DOI · Code · BibTeX

* Equal contribution.

Motivation

As AI systems become part of high-stakes workflows, users need evidence about which model produced a result and how it ran before they can trust it. That evidence should not reveal sensitive data. Verifiable AI sits at this boundary between systems, security, and machine learning.

verifiable.pro

SEPT Lab maintains verifiable.pro, a public resource on verifiable AI. It publishes news, explainers, and guides on what evidence can establish about an AI service’s claims, starting from trusted execution environments. Its verification stack places each kind of claim in one of seven layers, from the chip up to the laws, standards, and agreements that govern an AI system.