MedMisBench

Measuring Epistemic Resilience of LLMs Under Misleading Medical Context

MedMisBench evaluates whether large language models preserve the correct medical judgment when targeted misleading clinical context is introduced into a multiple-choice task. The benchmark covers five medical QA sources, five content-corruption types, and three provenance framings such as neutral false statements, patient claims, and authority-framed misinformation.

This folder contains simple, standalone notebooks for reproducing the core evaluation workflow.

Contents

File	Purpose
medmisbench_no_harness_eval.ipynb	Direct model evaluation with no search or browsing tools.
medmisbench_tool_harness_eval.ipynb	Tool-using evaluation with web search through Serper and page reading through Jina Reader.
medmisbench_overview.png	Overview figure used in this README.

Dataset

The dataset is hosted on Hugging Face:

from datasets import load_dataset

rows = load_dataset("HongjianZhou/MedMisBench", "MEDMISQA")["MEDMISQA"]
print(rows[0])

Released subsets:

Subset	Items	Role
MEDMISQA	3,112	Medical reasoning
MEDMISMCQA	3,986	Medical reasoning
MEDMISXPERTQA	1,544	Expert medical reasoning
MEDMISJOURNEY	2,197	Patient-journey evaluation
MEDMISHLE	93	Agentic biomedical capability

Evaluation Modes

Each notebook runs the same three settings:

Mode	Description
baseline	Original question and options only.
type1	Adds one targeted misleading context statement for a sampled wrong option.
type2	Adds all option-level context statements together, including support for the correct option and misleading context for wrong options.

Quick Start

1. Open one of the notebooks in Colab or Jupyter.
2. Run Step 1 to install dependencies.
3. In Step 2, paste the required API keys and set PROVIDER plus MODEL_ID.
4. For a smoke test, use:

HF_SUBSETS = ["MEDMISQA"]
END = 1

5. For a full benchmark run, keep all five subsets and set:

END = None

6. Run the notebook from top to bottom.

Supported Models

The no-harness notebook supports:

• OpenAI models through the OpenAI API
• Gemini models through google-genai
• Claude models through the Anthropic API
• local Hugging Face models served through an OpenAI-compatible SGLang endpoint

The tool-harness notebook supports OpenAI, Gemini, and Claude models, plus:

• SERPER_API_KEY for search_web
• JINA_API_KEY for visit_web

Output Layout

Each run writes a model-specific output folder:

medmisbench_no_harness_outputs/
  provider__model_id/
    MEDMISQA/
      baseline/
        raw/
        evaluations/
        raw_outputs.jsonl
        evaluations.jsonl
        results.csv
        summary.json
        run.log.txt

The tool-harness notebook uses the same layout under medmisbench_with_harness_outputs/ and additionally saves tool calls and visited evidence in the raw outputs.

Reading Results

The summary cell prints values in this format:

[baseline, type1, type2]

For example, a row such as [78.2, 45.1, 61.7] means:

• 78.2: accuracy on the original benchmark questions
• 45.1: accuracy after one targeted misleading context injection
• 61.7: accuracy when all option-level context statements are shown together

For detailed analysis, use evaluations.jsonl or results.csv. For prompt-level debugging, use the files under raw/ and the readable run.log.txt.

Safety Notice

MedMisBench intentionally contains false and misleading medical statements. It is intended for robustness evaluation and mitigation research, not clinical guidance or supervised training without safeguards.

Citation

If you use MedMisBench, please cite the manuscript.

@misc{zhou2026measuringepistemicresiliencellms,
      title={Measuring Epistemic Resilience of LLMs Under Misleading Medical Context}, 
      author={Hongjian Zhou and Xinyu Zou and Jinge Wu and Sean Wu and Junchi Yu and Bradley Max Segal and Tobias Erich Niebuhr and Sara Amro and Michael Petrus and Sheikh Momin and Alexandra M. Cardoso Pinto and Rachel Niesen and Laura Sophie Wegner and Dhruv Darji and Jung Moses Koo and Joshua Fieggen and Kapil Narain and Mingde Zeng and Lei Clifton and Linda Shapiro and Fenglin Liu and David A. Clifton},
      year={2026},
      eprint={2606.12291},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2606.12291}, 
}