Paperdoi:10.5281/zenodo.23570135
Do multimodal language models defer to a clinician's wrong ECG read? A controlled study of sycophancy in 12-lead ECG interpretation
Preprint/draft v0.8 in repo, targeting Journal of Electrocardiology; Zenodo software record v1.0.0 (2026-10-22)
Open code
TL;DR
Setup: 101 PTB-XL ECGs in 8 groups (25 acute-stage and 10 old infarctions) were rendered as paper images. Eight MLLMs read each case 3×. Cases a model got right all three times were re-shown with a clinician-style wrong hint (3 styles) or a correct hint, in Thai and in English. An open-ended sub-study added casual remarks.
- Form factor: clinical
Abstract
Code, prompts, case list, rendered ECG images (derived from PTB-XL, CC BY 4.0), raw responses of eight multimodal large language models and analysis scripts for a controlled study of sycophancy in 12-lead ECG interpretation. Code is MIT-licensed; data are CC BY 4.0.
Verbatim from DataCite (opens in a new tab) · Licence: https://opensource.org/licenses/MIT
Results on ChatPPG leaderboards
All leaderboardsNo results from this paper on ChatPPG leaderboards yet. Browse the leaderboards.
Tasks & signals
Signals
Keywords
Datasets
How it works
Cite this paper
APA
Do multimodal language models defer to a clinician's wrong ECG read? A controlled study of sycophancy in 12-lead ECG interpretation. (2026). Preprint/draft v0.8 in repo, targeting Journal of Electrocardiology. https://doi.org/10.5281/zenodo.23570135
BibTeX
@misc{sittikongmllmecg2026multimodal,
title = {{Do multimodal language models defer to a clinician's wrong ECG read? A controlled study of sycophancy in 12-lead ECG interpretation}},
year = {2026},
howpublished = {Preprint/draft v0.8 in repo, targeting Journal of Electrocardiology},
doi = {10.5281/zenodo.23570135}
}Built from the hub record. Check it against the publisher before submitting.
Added to the ChatPPG research hub
