Artifacts or Abduction: How Do LLMs Answer Multiple-Choice Questions Without the Question? — part 1/4

Cleaned-text code points [0, 15689). Bibliography omitted here; source.md includes it.

# Artifacts or Abduction: How Do LLMs Answer Multiple-Choice Questions Without the Question?

###### 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. Inferring the original question is an impressive reasoning strategy, but it cannot fully explain the high choices-only accuracy of LLMs in MCQA. Thus, while LLMs are not fully incapable of reasoning in MCQA, we still advocate for the use of stronger baselines in MCQA benchmarks, the design of robust MCQA datasets for fair evaluations, and further efforts to explain LLM decision-making. 1 1 1 Our data and code are available at https://github.com/nbalepur/mcqa-artifacts

## 1 Introduction

Multiple-choice question answering (MCQA) has been key for evaluating large language models (LLMs), valued for its ease of scoring and alignment with human testing protocols Robinson and Wingate (2023) . In this task, a model is given a question and list of choices as input, and must select the choice that best answers the question. This process is intended to assess if LLMs comprehend, use relevant pretraining knowledge, and reason via information from both the question and the choices. Since MCQA is meant to evaluate several model abilities, it is often used to rank LLMs in benchmarks such as the Open LLM leaderboard Beeching et al. (2023) and the Holistic Evaluation of Language Models Liang et al. (2023) benchmark.

Given this role of MCQA, it is essential to verify that MCQA accuracy reflects the abilities we intend to measure. While this concern has been studied with prompt sensitivity Zheng et al. (2024) , an unexplored aspect of MCQA-based LLM evaluation is dataset artifacts—patterns or biases in text that models may use as shortcuts, rather than completing the task as intended. Researchers have found that partial-input baselines Poliak et al. (2018) , models trained using just a subset of inputs, can exploit artifacts in parts of the input. Such models include hypothesis -only models in natural language inference Poliak et al. (2018) ; Herlihy and Rudinger (2021) and passage -only models in reading comprehension Shah et al. (2020) ; Kaushik and Lipton (2018) . If LLMs rely on artifacts, our datasets may not only fail to assess the skills they are designed to, but we may also overestimate LLMs and face generalizability issues in deployment Wiegreffe and Marasović (2021) .

Thus, to gauge if MCQA measures LLM abilities as intended, we probe if LLaMA-2 Touvron et al. (2023) , Falcon Penedo et al. (2023) , Phi-2 Abdin et al. (2023) , and Mixtral Jiang et al. (2024) can use artifacts on three benchmarks: ARC Clark et al. (2018) , MMLU Hendrycks et al. (2021) , and HellaSwag Zellers et al. (2019) . To do so, we test a partial-input choices -only model in MCQA, which omits the question and only sees the choices. Concretely, we propose partial-input prompts , a strategy analogous to trainable partial-input models, where only a subset of inputs are used in a prompt.

For our tested LLMs and datasets, we find that in 11/12 cases, choices-only accuracies significantly surpass majority baselines, with a sizeable gain of 0.33 in LLaMA on HellaSwag. Notably, this is achieved with minimal few-shot examples, ranging from 5 on MMLU to 25 on ARC. To give more insights, we design tests to help explain how LLMs obtain high choices-only accuracy and see if this behavior stems from surface-level shortcuts usually associated with artifacts Du et al. (2023) , or if more advanced reasoning is employed. We test three hypotheses, shown in Figure 1 (right).

1) Memorization: One explanation is that the LLM has been trained on the test set and is recalling answers via memorization ( section 4 ). To test this, we create partial-input prompts that give no discriminative information in the choices, and can only be answered if the LLM has already seen the answer (Figure 1 , top ). These prompts never largely best the majority baseline, leading us to test more complex behaviors to explain choices-only accuracy.

2) Choice Dynamics: With no evidence of memorization, we define two properties of choices that could explain the high choices-only accuracy ( section 5 ). First, LLMs may have individual priors on choices, such as favoring choices that contain “always.” Second, LLMs may use the group dynamics of choices, where the models assess a choice based on the other choices, such as favoring an odd-numbered choice if the other choices are even. We refer to these latter behaviors as meta-strategies , where the model reasons over all choices beyond individual priors.

To quantify the impact of individual priors on choices-only accuracy, we prompt the LLM to classify the correctness of each choice in isolation (Figure 1 , middle ). This prompt only sees one choice at a time and thus the LLM can only use its individual priors. In several cases, the individual priors alone cannot explain the high choices-only accuracy, implying that LLMs reason over all choices rather than just using cues from individual choices.

3) Abductive Question Inference: Inspired by work using LLMs to verbalize inferences Hoyle et al. (2023) and given the potential use of meta-strategies, we see if LLMs can reason via the meta-strategy of abductive question inference (AQI) ( section 6 ). We design a two-step prompt where the LLM: 1) abductively infers the missing question from the choices; and 2) answers its own question (Figure 1 , bottom ). We call it abductive QI as we aim to generate the question that best explains the observed choices Peirce (1974) ; Bhagavatula et al. (2020) .

AQI performs similarly to the choices-only prompt, even besting it for 3 LLMs on HellaSwag, and they also exhibit moderate agreement, implying that AQI may be just one of many strategies LLMs use in choices-only settings. We then study the questions generated by LLaMA on ARC. Out of the 43 cases when AQI picks the gold answer and the inferred question is answerable by one of the choices, 42% match the meaning of the original question . Thus, we argue that a successful choices-only prompt does not always mean that the LLM uses surface-level shortcuts linked with artifacts; in some cases, it may recover the missing question.

Our intention is not to discredit MCQA evaluations. Instead, by highlighting the high choices-only accuracy of LLMs and investigating several potential explanations, we aim to give suggestions to enhance transparency and robustness in benchmarking, as well as provide insights into the decision-making of LLMs in partial-input settings. Our contributions can be summarized as follows:

1) We assess choices-only prompts with four LLMs and three MCQA datasets, being the first to show that LLMs can exhibit high partial-input accuracy in few-shot settings and with minimal exemplars. 2) We test several hypotheses to help explain this high choices-only accuracy, such as memorization, choice priors, and abductive question inference. 3) We release a thorough black-box MCQA analysis suite to facilitate transparent LLM evaluations.

## 2 Experimental Setup

### 2.1 MCQA Task Definition

Our MCQA task is based on Robinson and Wingate (2023) , where the model is given: 1) a question q q ; and 2) a set of four choices 𝒞 = { c a , c b , c c , c d } \mathcal{C}=\{c_{a},c_{b},c_{c},c_{d}\} , exactly one of which is correct (i.e. gold choice c g ∈ 𝒞 c_{g}\in\mathcal{C} ). Using these inputs, the LLM must give the letter of the correct option a ∈ { (A) , (B) , (C) , (D) } a\in\{\texttt{(A)},\texttt{(B)},\texttt{(C)},\texttt{(D)}\} .

The standard MCQA prompt is a full prompt , where the LLM uses the question q q and choices 𝒞 \mathcal{C} :

In the prompt box above, and all prompt boxes in this paper, the non-highlighted text represents the input prompt provided to the model, while the highlighted text represents what the model generates (e.g. the letter of the choice in Prompt 2.1 ). All of our prompts are few-shot to guide the LLM outputs, leaving zero-shot artifact exploitation for future work. In the few-shot prompts, each exemplar follows the same format shown in the prompt box. For the exemplars, the highlighted text is replaced with the ground truth (Example in Appendix A.4 ).

### 2.2 Models

Due to the cost of closed-source LLMs, we study four open-source LLMs: LLaMA-2 70B 2 2 2 LLaMA is 8-bit quantized on MMLU to use less memory. Touvron et al. (2023) , Falcon 40B 3 3 3 Falcon is 20-shot on ARC to fit its 2048 token limit. Penedo et al. (2023) , Mixtral 8x7B Jiang et al. (2024) , and Phi-2 (2.7B) Abdin et al. (2023) . Each LLM is run with HuggingFace. We generate with default parameters and set the output length between 5 and 200 tokens.

### 2.3 Datasets

We select three MCQA benchmarks from the Open LLM Leaderboard Beeching et al. (2023) : ARC -Challenge Clark et al. (2018) , MMLU Hendrycks et al. (2021) , and HellaSwag Zellers et al. (2019) , collectively testing factual knowledge, scientific reasoning, and commonsense reasoning. We use the entire evaluation sets, but omit the questions in ARC that do not have four choices to ensure all questions have four choices. We follow the Open LLM Leaderboard to pick few-shot examples and use 25 random training examples for ARC, 10 random training examples for HellaSwag, and the 5 given training examples for each subject in MMLU. While our experiments reflect the prompting setups of present MCQA benchmarks, we believe that future works could extend our analysis with zero-shot prompting Kojima et al. (2022) to limit the influence of priors from the few-shot examples.

### 2.4 Evaluation Protocol

As defined in section 2.1 , the LLM aims to generate the letter of the gold answer choice. This black-box setup allows us to study LLM behavior without accessing LLM internals. However, one drawback is that LLMs may produce invalid outputs (e.g. “(E)” when there are four choices), especially when using atypical prompts like the memorization prompts in Figure 1 (top). In these cases, the LLM may not understand the task even with few-shot examples.

One solution is to mark every invalid output as incorrect, but this unfairly penalizes the LLM when the mistake likely stems from the atypical prompt. Thus, for a fair evaluation, when the LLM gives an invalid output, we treat it as random guessing and assign a score of 0.25. In Appendix A.5 , we show versions of all experiments where invalid outputs are marked as wrong, which do not alter our claims.

## 3 Performing MCQA With No Question

To test if LLMs exploit MCQA artifacts, we design partial-input choices-only prompts ( section 3.1 ), and show that these prompts often largely best majority baselines 4 4 4 A majority class baseline always predicts the most frequent answer choice found in the dataset. ( section 3.2 ). High partial-input accuracies typically imply artifact exploitation via simple shortcuts Du et al. (2023) . However, in § 6 we explore if some of this overperformance can be attributed to the LLM’s ability to infer the original question. In this case, it would be difficult to claim that the model has bypassed the intended MCQA format. Thus, in this section, we establish an initial ceiling for artifact exploitation in the choices-only setting.

### 3.1 Prompt Design

Artifacts can be uncovered via partial-input models Poliak et al. (2018) —trained models that omit parts of the input. To adapt this for LLMs, we propose an analogous method: partial-input prompts . In MCQA, the apt partial-input prompt is a choices-only prompt, where LLMs only use the choices:

We expect Prompt 3.1 to perform near a majority baseline, as ignoring the question should ideally equal random guessing. An accuracy largely above this suggests that the model may use artifacts.

### 3.2 Results

In 11/12 cases, the choices-only prompt surpasses the majority baseline significantly, indicating that LLMs may be using artifacts in MCQA benchmarks (Figure 2 ). Further, larger LLMs tend to have higher choices-only accuracy, implying that more capable LLMs may use artifacts to a higher degree. In Appendix B.1 , we study scaling laws with LLaMA to understand this relation more. Overall, these results serve as an initial ceiling for artifact exploitation in choices-only settings, which we use to motivate our subsequent analyses that help attribute where this accuracy may stem from.

Takeaway: Prior work has found artifacts with trained partial-input models Gururangan et al. (2018) , but we are the first to show that LLMs exhibit high choices-only accuracy in MCQA benchmarks, even in few-shot settings with limited exemplars. For researchers seeking to have LLMs perform MCQA as intended, where both the question and choices are needed, we recommend three approaches: First, along with typical MCQA metrics, choices-only prompts can be reported as stronger alternatives to majority baselines Poliak et al. (2018) . Second, researchers can design datasets with more robust protocols, such as the Winograd pair format Levesque et al. (2011) , to mitigate the potential for artifacts. Third, HellaSwag has the highest choices-only accuracy (0.585 with LLaMA) and it is also the only dataset with human -written gold answers and model -written distractors. Thus, the gold answers may contain stylistic cues distinct from the distractors, which can inform strong discriminators like LLMs in partial-input settings. To avoid introducing such artifacts, we advise researchers to use a consistent approach when generating text data.

## 4 Hypothesis 1: Memorization

Our first hypothesis to explain choices-only accuracy is test set leakage, where the LLM is trained on the test set Zhou et al. (2023) . If this occurred, the LLM could recall the answer via memorization Huang et al. (2022) . To test this, we design prompts only answerable via memorization ( section 4.1 ). We assess these prompts and find that memorization cannot explain choices-only accuracy ( section 4.2 ).

### 4.1 Prompt Design

While it is possible to test memorization via contamination analysis Sainz et al. (2023) , this does not reveal how memorization affects MCQA accuracy. As a simple solution, we create partial-input prompts where the LLM must return the letter a a of the correct choice, but without any discriminative information in the choices. Such prompts are only answerable if the LLM has already been trained on the example, allowing us to quantify how memorization alone impacts choices-only accuracy:

We cannot detect all forms of memorization, as adversarial actors could train an LLM on the test set while shuffling the choice order, bypassing our prompts. However, we assume a non-adversarial setting to test exact memorization and conjecture that if substantial exact memorization had occurred, these prompts would best the majority baseline.

### 4.2 Results