Reproduce “Position, Not Provenance: Separating Reasoning Mediation from Sycophancy in Medical Vision-Language Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Supratik Bhowal, Subhrajyoti Basu, Aritra Gir Mahanta, Anik Pal Chowdhury
CoT-Mediate audits whether edited reasoning changes medical vision-language model answers by altering one clinically meaningful detail in a model’s explanation and testing it as evidence or preceding context. On closed VQA-RAD questions, LLaVA-Med’s mediation score was 33.3% with evidence and 73.5% with prefix-forced continuation. The audit covers two models and measures prompt behavior, not internal reasoning or clinical accuracy.
Key insights
- Reasoning faithfulness in medical VLMs is not an inherent model property but is highly dependent on the method of reasoning injection, with prefix-forced continuation demonstrating stronger causal mediation than re-prompting.
- Models exhibit distinct behaviors regarding provenance: LLaVA-Med is authority-blind but self-critical, scrutinizing its own asserted reasoning more closely, while MedGemma shows instructed sycophancy and an authority gradient, discounting lower-authority sources.
- Laterality (e.g., left vs. right) was consistently the least tracked clinically load-bearing attribute across models and injection methods, indicating a specific weakness that might be overlooked by aggregate faithfulness scores.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.