CLR-Bench: Evaluating Large Language Models in College-level Reasoning

TL;DR

The results demonstrate the key insights that LLMs, even the best closed-source LLM, i.e., GPT-4 turbo, tend to `guess' the college-level answers.

Abstract

Large language models (LLMs) have demonstrated their remarkable performance across various language understanding tasks. While emerging benchmarks have been proposed to evaluate LLMs in various domains such as mathematics and computer science, they merely measure the accuracy in terms of the final prediction on multi-choice questions. However, it remains insufficient to verify the essential understanding of LLMs given a chosen choice. To fill this gap, we present CLR-Bench to comprehensively evaluate the LLMs in complex college-level reasoning. Specifically, (i) we prioritize 16 challenging college disciplines in computer science and artificial intelligence. The dataset contains 5 types of questions, while each question is associated with detailed explanations from experts. (ii) To quantify a fair evaluation of LLMs' reasoning ability, we formalize the criteria with two novel metrics. Q$\rightarrow$A is utilized to measure the performance of direct answer prediction, and Q$\rightarrow$AR effectively considers the joint ability to answer the question and provide rationale simultaneously. Extensive experiments are conducted with 40 LLMs over 1,018 discipline-specific questions. The results demonstrate the key insights that LLMs, even the best closed-source LLM, i.e., GPT-4 turbo, tend to `guess' the college-level answers. It shows a dramatic decrease in accuracy from 63.31% Q$\rightarrow$A to 39.00% Q$\rightarrow$AR, indicating an unsatisfactory reasoning ability.

Figures (6)

• Figure 1: (a) illustrates a sketched overview of the current pipeline to benchmark LLMs with multi-choice questions, while our CLR-Bench proposes a more comprehensive one with a variety of question types and verifies the college-level reasoning ability of LLMs as shown in (b) and (c).
• Figure 2: A real question answered by Qwen 2.5 and GPT-3.5. While GPT-3.5 has successfully identified the correct choices, it made the wrong rationale indicating the potential 'guessing' behavior.
• Figure 3: The overview of our proposed CLR-Bench. Dataset Construction. Domain experts first curate a condensed hierarchical topic graph to guide the collection of five types of questions. GPT-4o is then carefully instructed to assist the experts in gold rationale generation. Benchmark Evaluation. We formally define standardized criteria for each type of question and the corresponding rationale.
• Figure 4: A discipline-level model performance in terms of both Q$\rightarrow$A and Q$\rightarrow$AR. We selectively choose Qwen 2.5 72b, GPT-4 turbo and Claude3-opus to showcase their mastery of the reasoning ability across different topics.
• Figure 5: We selectively show the performance among closed- and open-source LLMs as well as within each group of models. The comparisons are made based on Q$\rightarrow$A, Q$\rightarrow$R and Q$\rightarrow$AR.