Property Finder Fixes Evaluation Bias and Optimizes Prompts for Model Selection

While performing model evaluation and prompt optimization on Amazon Bedrock, Property Finder discovered a key problem: 96.8% of its evaluation labels were inherited directly from predictions of the existing VLM rather than human verification, causing any alternative model to be inherently disadvantaged in comparisons and biasing the evaluation results. After discovering the issue, the team corrected the evaluation labels and resumed model evaluation and prompt optimization work on Amazon Bedrock. After fixing the evaluation bias, model accuracy improved while costs dropped by 86%. The whole story was reported by the AWS Machine Learning Blog on October 28, 2026, as a first-hand disclosure. The latest disclosed state of the event: by fixing evaluation bias and optimizing prompts for model selection, Property Finder achieved improved accuracy and an 86% cost reduction.

AI overview of reporting · Updated 2 h前

Not enough continuous observations to show a trend yet.