Research
Commit-Open Protocol Detects Silent Model Substitution in Hosted LLM Services Using SAE Feature Traces
Proposes a Merkle-tree-based commit-open protocol using sparse autoencoder (SAE) feature traces to detect when hosted LLM providers silently substitute a cheaper model for an advertised stronger one. Prior verification schemes (like SVIP) leave a parallel-serve side-channel exploitable by dishonest providers. This is the first cryptographic approach to closing the model-substitution gap — relevant as more teams pay premium pricing for specific model tiers.
Source
↳ Follow the thread