MUSE-AI-Lab/ReMCTS

Wang Huifei1 · Huang Xinying2 · Sun Yiheng1 · Yifan Yuan2† 1College of Computer Science and Software Engineering, Shenzhen University 2School of Artificial Intelligence, Shenzhen University †Corresponding author EMNLP 2026 ReMCTS: selection, expansion, execution, reflection and backpropagation over a tree of candidate programs. TL;DR: LLM-written code often looks right but fails on edge cases and hidden tests, and single-trajectory self-repair keeps patching an early wrong idea. ReMCTS searches over a tree of candidate programs: every candidate is executed in a sandbox, failures become structured reflections that the branch keeps (path memory) and that other branches retrieve (task-local global memory), and an availability-aware evidence reward tells failed checks apart from missing evidence, so the search also works when no tests are given. On HumanEval and MBPP-Sanitized, ReMCTS with visible tests beats direct generation in 8 of 10 model–dataset pairs; controlled sampling, repair, tree-search and memory ablations show where the gains come…