Determine whether prior VQA-RAD exposure affected Qwen-model results

Determine whether the evaluated Qwen2.5-VL-3B, Qwen2.5-VL-7B, and Qwen3-VL-8B models had prior exposure to the VQA-RAD items used in the evaluation, and thereby assess whether such exposure influenced the reported findings.

Background

The evaluation samples questions from the VQA-RAD training split, and the paper notes that the medical MedGemma-4B model was trained on VQA-RAD and may therefore have encountered some of the evaluation items. The authors separately state that prior exposure by the Qwen models is unknown, leaving unresolved whether memorization or prior familiarity could have affected their performance and sycophancy measurements.

References

Prior exposure of the Qwen models is unknown.

— Seeing Through the Forgery Is Not Enough: Regressive Sycophancy under Fabricated Evidence in Medical Vision-Language Models  (2610.17989 - Ahmed et al., 13 Oct 2026) in Limitations, paragraph “Models and data”