mahzzz/Sycophancy-Dataset

Talked Out of the Truth: Sycophancy in the Reasoning Chains of Multimodal Models

A benchmark for measuring sycophancy in large multimodal reasoning models, where a user asserts a wrong answer and the model may abandon correct visual reasoning under pressure. Each item pairs a visually grounded question with a constructed wrong answer, used to probe the model in single-turn and multi-turn pressure settings. Sycophancy is scored both in the final answer and within the reasoning chain.

Dataset Structure

462 items drawn from four source datasets (ClockQA 62, PathVQA 150, MathVision 150, SB-Bench 100). Each row has:

• dataset: source dataset name (clockqa, pathvqa, mathvision, sbbench)
• sample_idx: index within the source dataset
• question: the question text
• options: newline-separated multiple-choice options (empty for ClockQA and PathVQA)
• correct_answer: the ground-truth answer
• wrong_answer: the constructed wrong answer asserted to the model under pressure
• image: the image for the item

The wrong answer is constructed per dataset: a perturbed time for ClockQA, the inverted binary response for PathVQA, and an incorrect multiple-choice option for MathVision and SB-Bench.

Sources and Licenses

This dataset is built from four public datasets, and each item retains the terms of its source:

• PathVQA (He et al., 2020), MIT License — https://huggingface.co/datasets/flaviagiammarino/path-vqa
• MATH-Vision (Wang et al., 2024), MIT License — https://huggingface.co/datasets/MathLLMs/MathVision
• Lost in Time / DateTimeQA (Saxena et al., 2025), CC BY 4.0 — https://huggingface.co/datasets/rohitsaxena/DateTimeQA
• SB-Bench / BBQ-V (Narnaware et al., 2025), CC BY-NC 4.0 — https://huggingface.co/datasets/ucf-crcv/BBQ-V

License and Intended Use

This dataset is released under CC BY-NC 4.0. Use is non-commercial and research-only, consistent with the most restrictive source license (SB-Bench, CC BY-NC 4.0). Please attribute this dataset and the four source datasets above.

Content Warning

The SB-Bench portion contains contexts and questions that surface harmful social stereotypes. It is included solely for measuring and mitigating bias in AI systems.

Citation

If you use this dataset, please cite our paper and the four source datasets.

@article{islam2026talked, title={Talked Out of the Truth: Sycophancy in the Reasoning Chains of Multimodal Models}, author={{Numayeer Islam}, Mahir and Okuyama, Gakuto and Siauw, Nikolaus and Garg, Shivank and Panwar, Madhur and Sharma, Vasu}, journal={arXiv preprint arXiv:2608.28623}, year={2026} }