Model Equality Testing: Which Model Is This API Serving? (arxiv.org)

A recent study has introduced a concept called Model Equality Testing, which addresses the challenge of identifying changes in large language models accessed through APIs. As users increasingly rely on these black-box systems—such as Llama models via platforms like Amazon Bedrock and Azure AI Studio—there is a risk that model providers may alter their products through quantization, watermarking, or fine-tuning without informing users. The study formalizes the detection of these alterations through a statistical testing framework, allowing users to compare the distributions of outputs from these APIs against a reference distribution. The findings are significant for the AI/ML community as they establish a robust method for users to verify model integrity and performance, ensuring transparency and consistency. Using the Maximum Mean Discrepancy metric, the proposed test demonstrated a median power of 77.4% in detecting distortions, requiring an average of just 10 samples per prompt. When applied to commercial inference APIs for several Llama models, researchers revealed that 11 out of 31 endpoints produced outputs from different distributions than the original weights released by Meta, highlighting potential discrepancies that could impact application performance and user trust.