SWE-Explore: Benchmark for Coding Agent Exploration

Описание: In this AI Research Roundup episode, Alex discusses the paper: 'SWE-Explore: Benchmarking How Coding Agents Explore Repositories' Holistic benchmarks like SWE-bench often conflate exploration, bug localization, and patch generation, making it difficult to isolate why coding agents fail. To solve this, the authors introduce SWE-Explore, a new benchmark that isolates repository exploration as a ranked, line-level context-selection task. The benchmark covers 848 issues across 10 programming languages and 203 repositories, using a trajectory-grounded approach to establish ground-truth code regions. Evaluation of these explorers proves that upstream metrics like context efficiency and recall strongly track downstream patch success. Ultimately, SWE-Explore provides a fine-grained evaluation framework to understand and improve how LLM agents navigate complex codebases. Paper URL: https://arxiv.org/abs/2606.07297 #AI #MachineLearning #DeepLearning #CodingAgents #SoftwareEngineering #LLMs #SWEbench

Resources:
GitHub: https://github.com/Qiushao-E/SWE-Expl...