I gave an invited talk on Efficient Zero-Knowledge Proofs for AI Inference at the AViD (Assurance & Verification of AI Development) Workshop, organized by FAR.AI, Center for AI Safety, and Longview Philanthropy on May 17, 2026.

Speaking at AViD Workshop 2026


As AI moves into high-stakes domains like medical diagnosis, we need verifiable evidence that AI services actually ran the claimed model. Zero-knowledge machine learning (zkML) addresses this by using cryptographic proofs to certify correct AI inference while keeping model weights confidential. However, the prover overhead has been a major bottleneck (>1000x compared to plaintext inference).

In this talk, I presented our recent work that achieves a 10x improvement over the state of the art, reducing GPT-2 per-token prover time to 1.5 seconds (down from 3600s just three years ago). Our key techniques include:

  • A compiler that unifies diverse linear operations (MatMul, attention, etc.) into a single ZKP constraint
  • Sublinear-time proving (in the lookup table size) for exponentiation-related nonlinearities (e.g., softmax) by exploiting the special structure of lookup tables
  • Systems optimizations including batching and parallel proof computation

The full playlist of talks is available on YouTube.