MedProb
A lightweight probing framework that addresses medical visual question answering by extracting signals directly from the internal representations of a frozen vision-language model, bypassing the free-text generation stage. It achieves strong performance on multiple-choice Med-VQA benchmarks.
Timeline (last 90 days)
MedProb is a lightweight probing framework that bypasses the free-text generation step by directly predicting multiple-choice medical visual question answering answers from frozen vision-language model representations.
MedProb demonstrates significant advantages on three medical visual question answering benchmarks: PATH-VQA, SLAKE, and VQA-RAD.
Compared to conventional prompting methods, MedProb recovers more answer-relevant signals and surpasses both VLMs trained specifically for the medical domain and agent systems.
The research team conducted a systematic comparison of 14 matched pairs of general VLMs and medical VLMs, finding that medical domain-specific adaptation did not consistently improve linear decodability.
MedProb itself also exhibits position bias, but the patterns it is affected by are distinctly different from those of prompting methods.
The MedProb probing method can be extended to open-ended generation tasks through a rejection sampling scoring procedure.
The study challenges the conventional assumption that medical visual question answering requires medical domain-specific training of vision-language models.
The research suggests that the medical question-answering capabilities embedded within pre-trained models may be far richer than their generative outputs indicate.
All Facts (8)
MedProb is a lightweight probing framework that bypasses the free-text generation step by directly predicting multiple-choice medical visual question answering answers from frozen vision-language model representations.
50%UnverifiedMedProb demonstrates significant advantages on three medical visual question answering benchmarks: PATH-VQA, SLAKE, and VQA-RAD.
50%UnverifiedCompared to conventional prompting methods, MedProb recovers more answer-relevant signals and surpasses both VLMs trained specifically for the medical domain and agent systems.
50%UnverifiedThe research team conducted a systematic comparison of 14 matched pairs of general VLMs and medical VLMs, finding that medical domain-specific adaptation did not consistently improve linear decodability.
50%UnverifiedMedProb itself also exhibits position bias, but the patterns it is affected by are distinctly different from those of prompting methods.
50%UnverifiedThe MedProb probing method can be extended to open-ended generation tasks through a rejection sampling scoring procedure.
50%UnverifiedThe study challenges the conventional assumption that medical visual question answering requires medical domain-specific training of vision-language models.
50%UnverifiedThe research suggests that the medical question-answering capabilities embedded within pre-trained models may be far richer than their generative outputs indicate.
50%