[Paper Review] Multiple-Choice Questions are Efficient and Robust LLM Evaluators

Ziyin Zhang, Zhaokun Jiang+3 more|arXiv (Cornell University)|May 20, 2024

Customer churn and segmentationBusiness, Management and Accounting3 citations

TL;DR

This paper proposes converting open-ended LLM evaluation benchmarks like GSM8K and MATH into multiple-choice (MC) formats—GSM-MC and MATH-MC—by collecting correct answers and incorrect predictions from over 50 open-source models to create robust distractors. The method enables efficient, accurate evaluation via model logits, reducing evaluation time by up to 30x while maintaining strong correlation (r > 0.9) with original benchmark performance, even under varying distractor choices and option orders.

ABSTRACT

We present GSM-MC, a multiple-choice (MC) dataset constructed by collecting answers and incorrect predictions on GSM8K from 60 open-source models. Through extensive experiments, we show that LLMs' performance on the MC version of this popular benchmark is strongly correlated with their performance on the original version and is quite robust to distractor choices and option orders, while the evaluation time is reduced by a factor of up to 30. Following similar procedures, we introduce MATH-MC, constructed from MATH, and PythonIO, a new program reasoning MC dataset constructed from HumanEval and MBPP. Experimental results indicate that LLMs' performance on these MC benchmarks leaves much room for improvement. Our data and code are available at https://github.com/Geralt-Targaryen/MC-Evaluation.

Motivation & Objective

• Address the problem of invalid or format-incorrect answers in LLM evaluations on open-ended benchmarks like GSM8K and MATH.
• Reduce evaluation time and false negatives caused by heuristic answer extraction from model generations.
• Improve robustness of LLM evaluation by leveraging multiple-choice format with curated distractors from model predictions.
• Extend the MC evaluation paradigm to code generation benchmarks by introducing a program output prediction task (PythonIO).
• Demonstrate that MC evaluation is a reliable, efficient, and robust alternative to standard open-ended evaluation.

Proposed method

• Construct GSM-MC and MATH-MC by extracting correct answers and incorrect predictions from over 50 open-source LLMs on GSM8K and MATH.
• Use model-generated incorrect answers as distractors to form multiple-choice options, preserving semantic plausibility.
• Reformat original problems into MC format with 4–8 options, ensuring consistent structure and answer format.
• Develop PythonIO by converting HumanEval and MBPP into program output prediction tasks, where models select the correct output from multiple choices.
• Evaluate LLMs on MC versions using logits to extract predictions directly, avoiding error-prone text parsing.
• Apply systematic ablation studies across varying numbers of choices, distractor sets, and option order permutations to test robustness.

Experimental results

Research questions

• RQ1 Can multiple-choice conversion of open-ended benchmarks like GSM8K and MATH preserve model performance correlation with original benchmarks?
• RQ2 How robust is MC evaluation to variations in distractor quality and option ordering?
• RQ3 To what extent does MC evaluation reduce evaluation time compared to standard open-ended generation?
• RQ4 Can the MC evaluation paradigm be effectively extended to code generation benchmarks like HumanEval and MBPP?
• RQ5 Does MC evaluation mitigate false negatives caused by format-incorrect or invalid model outputs?

Key findings

• LLM performance on GSM-MC and MATH-MC shows a strong correlation (r > 0.9) with performance on the original GSM8K and MATH benchmarks across all tested model sizes and settings.
• MC evaluation is robust to distractor variation and option ordering, with performance stability maintained even when distractors are replaced with predictions from different models.
• Evaluation time is reduced by up to 30x by extracting predictions directly from model logits instead of parsing generated text.
• The proposed PythonIO dataset enables efficient and accurate evaluation of code generation models using multiple-choice format, avoiding parsing errors from code generation.
• Instruction-tuned models exhibit significantly more invalid answers in open-ended format, but MC conversion mitigates this issue by standardizing output format.
• The method maintains high evaluation fidelity even with as few as 2 choices, demonstrating scalability and reliability across different MC configurations.