Coin-flip deference and a small reliability response: a partial replication in medical vision–language models

Background. Earlier work reported that medical vision–language models scarcely adjust their use of diagnostic advice to its stated reliability. We tested the decision-level consequences with a separate pipeline, new patients and generated answers. Methods. A frozen local protocol, not a preregistration, specified 400 NIH ChestX-ray14 patients, balanced for effusion and selected after excluding patients with ChestX-Det thumbnail matches. MedGemma-4B and Lingshu-7B answered with no advice or either binary report from a simulated tool at 3 stated symmetric reliabilities. The primary arm disclosed the simulated error mechanism; a metadata-only arm omitted it. Primary endpoints were the change in counterfactual decision reliance, ∆R = R(0.99) − R(0.50), and the loss of balanced correct decisions with coin-flip advice, H50. We used class-stratified patient bootstrap intervals. Results. In the primary arm, coin-flip reliance was 19.8 [16.3, 23.3]% for MedGemma and 34.5 [30.8, 38.0]% for Lingshu (95% intervals). The corresponding H50 values were 6.4 [4.4, 8.4] and 10.0 [7.4, 12.6] pp. Both fell below always following a 99% tool, by 13.4 and 8.1 pp. MedGemma's ∆R was -0.3 [-1.5, 0.8] pp; Lingshu's was positive, 3.7 [2.0, 5.7] pp. Both 90% intervals lay inside the prespecified ±10 pp margin, but Lingshu's detectable response challenges blanket reliability blindness. Its positive response was supported only in the mechanism-explicit arm. Conclusion. Deference to disclosed uninformative advice and under-use of highly reliable advice replicated. Blanket blindness did not replicate for Lingshu. Generated decisions establish neither probability-level blindness nor clinical readiness; NLP-mined labels, selected normal negatives and possible training exposure limit interpretation.

Authors

Institutions

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-19
DOI
https://doi.org/10.5281/zenodo.24373649
Primary Topic
Artificial Intelligence in Healthcare and Education
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Coin-flip deference and a small reliability response: a partial replication in medical vision–language models

S. M. Abdulla Al Mamun
Zenodo (CERN European Organization for Nuclear Research)
Artificial Intelligence in Healthcare and Education
preprint

Coin-flip deference and a small reliability response: a partial replication in medical vision–language models

S. M. Abdulla Al Mamun
preprint en

Abstract

Background. Earlier work reported that medical vision–language models scarcely adjust their use of diagnostic advice to its stated reliability. We tested the decision-level consequences with a separate pipeline, new patients and generated answers. Methods. A frozen local protocol, not a preregistration, specified 400 NIH ChestX-ray14 patients, balanced for effusion and selected after excluding patients with ChestX-Det thumbnail matches. MedGemma-4B and Lingshu-7B answered with no advice or either binary report from a simulated tool at 3 stated symmetric reliabilities. The primary arm disclosed the simulated error mechanism; a metadata-only arm omitted it. Primary endpoints were the change in counterfactual decision reliance, ∆R = R(0.99) − R(0.50), and the loss of balanced correct decisions with coin-flip advice, H50. We used class-stratified patient bootstrap intervals. Results. In the primary arm, coin-flip reliance was 19.8 [16.3, 23.3]% for MedGemma and 34.5 [30.8, 38.0]% for Lingshu (95% intervals). The corresponding H50 values were 6.4 [4.4, 8.4] and 10.0 [7.4, 12.6] pp. Both fell below always following a 99% tool, by 13.4 and 8.1 pp. MedGemma's ∆R was -0.3 [-1.5, 0.8] pp; Lingshu's was positive, 3.7 [2.0, 5.7] pp. Both 90% intervals lay inside the prespecified ±10 pp margin, but Lingshu's detectable response challenges blanket reliability blindness. Its positive response was supported only in the mechanism-explicit arm. Conclusion. Deference to disclosed uninformative advice and under-use of highly reliable advice replicated. Blanket blindness did not replicate for Lingshu. Generated decisions establish neither probability-level blindness nor clinical readiness; NLP-mined labels, selected normal negatives and possible training exposure limit interpretation.

Zenodo (CERN European Organization for Nuclear Research)
University of Manchester (GB)
Artificial Intelligence in Healthcare and Education
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon