SIMIT Improves VLM Performance via Test-Time Imaginary Training Data

SIMIT is a test-time framework where vision-language models generate their own query-specific training examples. By imagining similar problems and their solutions, the model improves its performance on unlabeled test queries without external supervision.