Reproduce “UserSWE: Evaluating Coding Agents in Multi-Turn Conversations with Dynamic Users” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Zhan-Yu Wu, Ji Luo, Alyssa Wu, Junjielong Xu, Richeng Xuan, Zhichao Hu, Chongming Gao, Xinting Hu
UserSWE evaluates repository-coding agents in conversations where simulated users reveal bounded information over turns and tests run after each response. Its changing user personas are calibrated from developer sessions. Among 11 agents, Claude Fable 5.1 led at one turn with 31.9%, while Claude Opus 5 had the highest Length-Discounted Pass Rate (LDPR), 0.640; rankings varied by turn budget in UserSWE.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.