Test-Time Hinting for Black-Box Vision-Language Models
Test-Time Hinting (TTH) improves Vision-Language Model performance by prepending a targeted natural language hint to the prompt before a single VLM call. This method identifies likely failure modes based on input and uses a trained generator to provide guidance, steering the model away from errors. It works effectively with closed-weight models via API access, offering cost-efficient, single-pass inference.