TOPLOC introduces a novel locality sensitive hashing scheme that enables trustless verification of large language model inference, solving a critical security challenge.

• Creates a compact fingerprint of model activations to detect unauthorized modifications
• Eliminates need to trust inference providers about model configuration
• Enables verification without compromising model privacy or performance
• Provides a practical security layer for enterprise AI deployment

This research addresses growing concerns about AI security and integrity as organizations increasingly rely on third-party inference providers for advanced language models.

TOPLOC: A Locality Sensitive Hashing Scheme for Trustless Verifiable Inference