Large language models (LLMs) have demonstrated strong performance in solving math problems, and there is growing research on evaluating their robustness. Unlike previous studies that create problem variants by adding perturbations to a single problem, this paper focuses on the interaction between problems. Specifically, we combine two original problems with a logical connection to get a new math problem, and measure the LLMs’ performance on it to evaluate its compositional generalization, which is an important and essential reasoning capability in human intelligence. The result of experiments that cover 14 different LLMs shows that even when the mathematical essence remains unchanged, a simple form of combination can significantly reduce the performance of LLMs, revealing the limitation of their generalization ability. Additionally, we propose an automated pipeline with 98.2% accuracy to assist in annotating datasets (1 manual, 2 synthetic). The extensive experiments conducted on these datasets further verify the conclusion and obtain some important findings. Finally, we analyze the impact of factors such as difficulty and length on LLMs’ performance, offering insights for future research.

Large language models (LLMs) have manifested strong ability to generate codes for productive activities. However, current benchmarks for code synthesis, such as HumanEval, MBPP, and DS-1000, are predominantly oriented towards introductory tasks on algorithm and data science, insufficiently satisfying challenging requirements prevalent in real-world coding. To fill this gap, we propose NaturalCodeBench (NCB), a challenging code benchmark designed to mirror the complexity and variety of scenarios in real coding tasks. NCB comprises 402 high-quality problems in Python and Java, meticulously selected from natural user queries from online coding services, covering 6 different domains. Noting the extraordinary difficulty in creating testing cases for real-world queries, we also introduce a semi-automated pipeline to enhance the efficiency of test case construction. Comparing with manual solutions, it achieves an efficiency increase of more than 4 times. Our systematic experiments on 39 LLMs find that performance gaps on NCB between models with close HumanEval scores could still be significant, indicating a lack of focus on practical code synthesis scenarios or over-specified optimization on HumanEval. On the other hand, even the best-performing GPT-4 is still far from satisfying on NCB. The evaluation toolkit and development set are available at https://github.com/THUDM/NaturalCodeBench.