Type: url
Original: https://arxiv.org/html/2406.01422v2
Status: generated
Summary
This FSE 2025 companion-proceedings paper from Tongyi Lab, Alibaba Group presents LingmaAgent, an automated issue-resolution method deployed in the TONGYI Lingma coding assistant that first builds a repository knowledge graph (a file/class/function hierarchy extended with a function call graph) and then explores it with a Monte Carlo tree search guided by LLM-scored rewards, before summarizing the findings and generating patches with AutoCodeRover-style search APIs. On SWE-bench Lite it reports 21.33% resolved with GPT-4 Turbo, an 18.5% relative improvement over SWE-agent, and on an in-house Alibaba Cloud dataset of 194 issues it reports 16.9% resolved automatically and 43.3% on a 30-issue subset with human-in-the-loop intervention.