Multiple Choice Question Answering Artifacts

This repository is the official implementation of "Artifacts or Abduction: How Do LLMs Answer Multiple-Choice Questions Without the Question?", which was accepted to ACL 2024 and can be viewed on Arxiv here.

Overview

This repository contains the code and dataset to run our experiments related to choices-only LLMs in MCQA, memorization, individual priors and choices dynamics, and the meta-strategy of abductive question inference

Setup

Python 3.10.0, pip 23.2.1, and conda 23.5.0 were used when running the code in this repository. A list of requirements can be found in requirements.txt, which can be installed through the following command:

pip install -r requirements.txt 

The most important files in this repository are as follows:

• /model/: Contains the code for running all experiments with LLMs
• /prompts/: Contains the prompts used in the independent full and indepdent choices-only prompt experiments
• /evaluation/: Contains the code for generating our plots
• /scripts/: Sample bash scripts to run our code

Model Usage

There are six relevant files:

• /model/run_hf.py: Code to run all possible prompt formats, except the second step of the inferring the question strategy and the random question prompt
• /model/run_hf_remote.py: Version of run_hf.py that works with models on the huggingface hub where you need to trust remote code (i.e. Phi)
• /model/extract_generated_questions.py: After running step one of the inferring the question strategy in the above two Python scripts, this will extract the questions generated by the LLM. This must be run before either of the last two Python scripts
• /model/extract_random_questions.py: This will sample a random set of questions from the test set. for the LLM to use. This must be run before either of the last two Python files
• /model/run_hf_question.py: Code to run the second step of the inferring the question strategy and the random question prompt
• /model/run_hf_question_remote.py: Version of run_hf_question.py that works with models on the huggingface hub where you need to trust remote code (i.e. Phi)

Samples bash scripts for run_hf.py, run_hf_remote.py, run_hf_question.py and run_hf_question_remote.py can be found in the /scripts/ folder. All of these files require the following arguments, which are detailed in the bash scripts:

• model_name: Nickname of the model
• model_name_hf: Huggingface model name
• experiments: List of experiments to run, identified by ProomptType enum
• datasets: List of datasets to run, identified by DatasetName enum
• partition: Partition of the dataset to run (e.g. full, first_half)
• hf_token: huggingface token (for downloading gated models
• load_in_8bit: load the model in 8bit? ("False" or "True")
• load_in_4bit: load the model in 4bit? ("False" or "True")
• use_20_fewshot: use a 20-shot prompt? (only applies to ARC, as it it normally 25-shot)
• res_dir: Directory pointing to the results folder
• prompt_dir: Directory pointing to the prompt folder
• cache_dir: Directory pointing to the cache folder for saving the downloaded models

In addition, run_hf_question.py and run_hf_question_remote.py have the extra parameter:

• use_random_question: Should the two-step question in "Inferring the Question" be a randomly-generated question or a model-generated question? ("True or "False")

To run extract_generated_questions.py and extract_random_questions.py, it is best to specify the following parameters at the top of the file:

• res_dir: Directory pointing to the results folder
• DATASETS: List of datasets to extract the questions from
• MODELS: Models to extract the questions from

Evaluation Usage

There are two relevant files:

• /evaluation/plot_accuracy.py: Create bar charts for all models except the independent classification prompts
• /evaluation/plot_accuracy_individual.py: Create bar charts for only the independent classification prompts

The files use the following parameters specified at the top of the file:

• res_prefix: Directory pointing to the results folder
• out_dir: Output directory to save the plot

The file plot_accuracy.py also has the following parameter:

• EXPERIMENTS: List of experiments to show in the bar chart

Citation

If you found our code/papers useful, you can cite us as follows:

@inproceedings{balepur-etal-2024-artifacts,
    title = "Artifacts or Abduction: How Do {LLM}s Answer Multiple-Choice Questions Without the Question?",
    author = "Balepur, Nishant  and
      Ravichander, Abhilasha  and
      Rudinger, Rachel",
    editor = "Ku, Lun-Wei  and
      Martins, Andre  and
      Srikumar, Vivek",
    booktitle = "Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
    month = aug,
    year = "2024",
    address = "Bangkok, Thailand",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2024.acl-long.555",
    doi = "10.18653/v1/2024.acl-long.555",
    pages = "10308--10330",
    abstract = "Multiple-choice question answering (MCQA) is often used to evaluate large language models (LLMs). To see if MCQA assesses LLMs as intended, we probe if LLMs can perform MCQA with choices-only prompts, where models must select the correct answer only from the choices. In three MCQA datasets and four LLMs, this prompt bests a majority baseline in 11/12 cases, with up to 0.33 accuracy gain. To help explain this behavior, we conduct an in-depth, black-box analysis on memorization, choice dynamics, and question inference. Our key findings are threefold. First, we find no evidence that the choices-only accuracy stems from memorization alone. Second, priors over individual choices do not fully explain choices-only accuracy, hinting that LLMs use the group dynamics of choices. Third, LLMs have some ability to infer a relevant question from choices, and surprisingly can sometimes even match the original question. We hope to motivate the use of stronger baselines in MCQA benchmarks, the design of robust MCQA datasets, and further efforts to explain LLM decision-making.",
}

@inproceedings{balepur-rudinger-2024-large,
    title = "Is Your Large Language Model Knowledgeable or a Choices-Only Cheater?",
    author = "Balepur, Nishant  and
      Rudinger, Rachel",
    editor = "Li, Sha  and
      Li, Manling  and
      Zhang, Michael JQ  and
      Choi, Eunsol  and
      Geva, Mor  and
      Hase, Peter  and
      Ji, Heng",
    booktitle = "Proceedings of the 1st Workshop on Towards Knowledgeable Language Models (KnowLLM 2024)",
    month = aug,
    year = "2024",
    address = "Bangkok, Thailand",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2024.knowllm-1.2",
    doi = "10.18653/v1/2024.knowllm-1.2",
    pages = "15--26",
    abstract = "Recent work shows that large language models (LLMs) can answer multiple-choice questions using only the choices, but does this mean that MCQA leaderboard rankings of LLMs are largely influenced by abilities in choices-only settings? To answer this, we use a contrast set that probes if LLMs over-rely on choices-only shortcuts in MCQA. While previous works build contrast sets via expensive human annotations or model-generated data which can be biased, we employ graph mining to extract contrast sets from existing MCQA datasets. We use our method on UnifiedQA, a group of six commonsense reasoning datasets with high choices-only accuracy, to build an 820-question contrast set. After validating our contrast set, we test 12 LLMs, finding that these models do not exhibit reliance on choice-only shortcuts when given both the question and choices. Thus, despite the susceptibility of MCQA to high choices-only accuracy, we argue that LLMs are not obtaining high ranks on MCQA leaderboards solely due to their ability to exploit choices-only shortcuts.",
}