This dataset holds the rendered memes of the meme revision strategy in the paper Revising Context, Shifting Simulated Stance: Auditing LLM-Based Stance Simulation in Online Discussions (Findings of EMNLP 2026).

Paper: arXiv:2606.06443
Code and released outputs: github.com/STARResearchLab/StanceShift

In the meme strategy, a meme replaces the last message from another user (n_i) in a Reddit conversation, and the stance simulator infers the target user's stance under this revised context. Each image in memes/ is the exact image that the stance simulators (GPT-5.2, Sonnet-4.6, Qwen3.5-Plus) received in the paper. The GitHub repository releases the meme texts and the simulator outputs but not the images. This dataset completes that release.

Folder	Topic	Images	Conversations	Templates per conversation
`memes/deepseek/`	DeepSeek	3,934	787	5
`memes/claude/`	Claude	2,674	538	5
`memes/llama/`	Llama	2,472	495	5
`memes/nuclear/`	Nuclear energy (Appendix G)	296	296	1
`memes/keto/`	Keto diet (Appendix G)	300	300	1
`memes/ubi/`	Universal basic income (Appendix G)	296	296	1
`memes/spacexipo/`	SpaceX IPO (Appendix G)	300	300	1
`white_memes/{deepseek,claude,llama}/`	r_white_meme ablation (Table 4)	750	150	5

Only the images that were scored in the paper are included. A few meme texts in the GitHub release have no image here because the image could not be generated. For the additional domains, the template of each conversation is listed in data/template_ids/{topic}.json of the GitHub repository.

white_memes/ holds the r_white_meme ablation: the same meme text drawn on a plain white canvas instead of the meme template, for the 50 conversations per topic listed in data/ablation_ids/ of the GitHub repository.

File layout

memes/{topic}/{conv_id}_{template_id}.png
memes/metadata.csv
white_memes/{topic}/{conv_id}_{template_id}.png
white_memes/metadata.csv

conv_id matches the conversation ids in the GitHub repository (data/skeletons/{topic}.jsonl, outputs/revisions/{topic}_meme.jsonl, outputs/stances/{simulator}/{topic}_meme.csv). template_id is one of 0, 1, 9, 3, 4, which are meme templates 1 to 5 of Appendix H. The template images are in meme_templates/ of the GitHub repository.

metadata.csv has one row per image with the columns file_name, topic, conv_id, template_id, paper_template (1 to 5, the numbering of Appendix H), width, height, file_size (bytes) and sha256.

Usage

The folder layout matches the default --meme_dir outputs/memes of the StanceShift pipeline. From the root of the GitHub repository:

pip install -U huggingface_hub
hf download STARResearchLab/StanceShift-memes --repo-type dataset --include "memes/*" --local-dir outputs/
python pipeline/prepare_prompt.py -t deepseek -s meme --revised   # stance prompts with the released memes

To download a single topic, use for example --include "memes/claude/*". The images can also be loaded with datasets:

from datasets import load_dataset
memes = load_dataset("STARResearchLab/StanceShift-memes", "memes", split="train")

How the memes were made

For each conversation and template, the revision model (gemini-3-flash-preview) wrote the meme text for the positions of the template. The text was then added to the template image with an OpenAI GPT Image model through image editing. The prompts are in Appendix I of the paper and in pipeline/ of the GitHub repository. The white memes were drawn locally with PIL. The images are kept as generated, including rendering errors such as misspellings or changed layouts.

License

The images are released under CC BY-NC 4.0 for research purposes. The memes are derived from third-party meme templates collected from ImgFlip. The template images are not covered by this license, and their use is subject to the rights of their owners.

Responsible use

The memes are controlled probes for auditing how LLM-based stance simulation responds to changes in conversational context. Do not use them to persuade or manipulate real users or to fabricate public opinion. The measured stance shifts are changes in LLM-simulated stance and are not evidence of opinion change in real users.

Citation

@inproceedings{zhang2026revising,
  title     = {Revising Context, Shifting Simulated Stance: Auditing {LLM}-Based Stance Simulation in Online Discussions},
  author    = {Zhang, Xinnong and Shan, Wanting and Lyu, Hanjia and Wei, Zhongyu and Luo, Jiebo},
  booktitle = {Findings of the Association for Computational Linguistics: EMNLP 2026},
  year      = {2026}
}