1、Trae Agent:An LLM-based Agent for SoftwareEngineering with Test-time ScalingTrae ResearchBeijing,Chinaopensourcemail.trae.aiAbstractSoftware issue resolution is a critical challenge in software engineering and hasgarnered increasing attention in recent years.With the rapid advancement oflarge langua
2、ge models(LLMs),substantial progress has been made in addressingreal-world software engineering tasks.Recent studies have introduced ensemblereasoning techniques to enhance the performance of LLM-based issue resolution.However,existing prompting-based methods still face limitations in effectivelyexp
3、loring large ensemble spaces and lack the capacity for repository-level un-derstanding,both of which constrain their overall effectiveness.In this paper,we proposeTrae Agent,the fi rst agent-based ensemble reasoning approach forrepository-level issue resolution.Trae Agentformulates our goal as an op
4、timalsolution search problem and addresses two key challenges,i.e.,large ensemblespaces and repository-level understanding,through modular agents for generation,pruning,and selection.We conduct extensive experiments using three leadingLLMs on the widely-adopted SWE-bench benchmark,comparingTrae Agen
5、tagainst four state-of-the-art ensemble reasoning techniques.Experimental resultsdemonstrate thatTrae Agentconsistently achieves superior performance,withan average improvement of 10.22%over all baselines in terms of Pass1.TraeAgenthas achieved first place on the SWE-bench Verified leaderboard,with
6、anotable Pass1 score of 75.20%.We are pleased to releaseTrae Agentas anopen-source project to support the research community,with all resources availableat https:/ issue resolution refers to the automated handling of newly reported bugs or feature re-quests during software development and maintenanc