Fool Your (Vision and) Language Model with Embarrassingly Simple Permutations

Authors: Yongshuo Zong, Tingyang Yu, Ruchika Chavhan, Bingchen Zhao, Timothy Hospedales

ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Specifically, we show empirically that popular models are vulnerable to adversarial permutation in answer sets for multiple-choice prompting, which is surprising as models should ideally be as invariant to prompt permutation as humans are.
Researcher Affiliation Academia 1University of Einburgh 2EPFL.
Pseudocode No The paper provides a mathematical equation for the attack but does not include any pseudocode or algorithm blocks.
Open Source Code Yes Code is available at https://github.com/ ys-zong/Foolyour VLLMs.
Open Datasets Yes All of the datasets we use are publicly available. Specifically, for LLMs, we utilize MMLU (Hendrycks et al., 2020), ARC challenge (ARC-c) (Clark et al., 2018), Bool Q (Clark et al., 2019), Sociali QA (Sap et al., 2019), and Med MCQA (Pal et al., 2022). For VLLMs, we use Science QA (Lu et al., 2022), A-OKVQA (Schwenk et al., 2022), MMBench (Liu et al., 2023c), and SEED-Bench (Li et al., 2023a).
Dataset Splits Yes Specifically, for LLMs, we utilize MMLU (Hendrycks et al., 2020), ARC challenge (ARC-c) (Clark et al., 2018), Bool Q (Clark et al., 2019), Sociali QA (Sap et al., 2019), and Med MCQA (Pal et al., 2022). For VLLMs, we use Science QA (Lu et al., 2022), A-OKVQA (Schwenk et al., 2022), MMBench (Liu et al., 2023c), and SEED-Bench (Li et al., 2023a). ... As many of the benchmarks do not provide a training set, we conduct two fine-tuning experiments using Llama2-7B on two datasets that do provide training sets: ARC-Challenge (Clark et al., 2018) and Med MCQA (Pal et al., 2022).
Hardware Specification Yes Experiments are conducted on A100-80GB GPUs.
Software Dependencies No The paper mentions accessing model weights from Hugging Face or official repositories but does not explicitly list specific software dependencies with version numbers (e.g., Python, PyTorch, TensorFlow versions).
Experiment Setup Yes For both LLMs and VLLMs, we use greedy decoding to ensure reproducibility. ... We fine-tune with Lo RA (Hu et al., 2022) for 1 epoch.