Search papers, labs, and topics across Lattice.
SWE-QA, a new benchmark, is introduced to evaluate multi-hop code comprehension by posing questions that require reasoning across multiple code segments within 12 Python repositories from SWE-bench. The dataset contains 9,072 multiple-choice questions generated using parsing-based entity extraction and LLM assistance, focusing on reasoning patterns like Declaration-and-Call and Interacting-Entity relationships. Evaluation of 15 LLMs (360M to 671B parameters) reveals that even the best models struggle with multi-hop reasoning, achieving only 74.41% accuracy, and dense architectures outperform MoE models.
Even the largest language models still struggle to connect information across dispersed code segments, achieving only 74% accuracy on a new benchmark designed to test multi-hop code comprehension.
In this paper, we introduce SWE-QA, a text and code corpus aimed at benchmarking multi-hop code comprehension, addressing the gap between simplified evaluation tasks and the complex reasoning required in real-world software development. While existing code understanding benchmarks focus on isolated snippets, developers must routinely connect information across multiple dispersed code segments. The dataset comprises 9,072 multiple-choice questions systematically generated from 12 Python repositories of SWE-bench, evaluating several recurrent reasoning patterns like Declaration-and-Call questions that link entity definitions to their usage, and Interacting-Entity questions that examine the dynamic relationships among multiple collaborating components. Generated through parsing-based entity extraction and Large Language Model assisted question construction with carefully validated distractors, the benchmark distinguishes genuine comprehension from superficial pattern matching. Evaluation of 15 language models (360M to 671B parameters) reveals significant challenges in multi-hop reasoning, with best performance reaching 74.41% accuracy. Dense architectures consistently outperform mixture-of-experts models by 10-14 percentage points, while reasoning-enhanced variants show inconsistent benefits.
