Memdoor: A team of 5 Qwen3.8-27B models beats Fable 5 on LiveCodeBench Hard (github.com)

In a significant advancement for AI programming solutions, a team utilizing five Qwen3.8-27B models surpassed the performance of Fable 5 by achieving a mean score of 9.8 out of 10 on the challenging LiveCodeBench Hard problems. This accomplishment was facilitated by a multi-agent workflow called Memdoor, which structured the collaborative problem-solving process through distinct roles (planner, ideators, solver, manager, and repair worker), all communicating via files on disk. The framework emphasizes consistency across model prompts and workflows, with results generated through a shell script rather than dependency on the models alone. The implications of this success are profound. It underscores the potential for structured multi-agent collaboration in tackling complex programming challenges, enhancing both efficiency and accuracy. The integration of rigorous fiscal details—such as the $35.79 cost for processing the first 50 problems—and the transparent reporting on model performance adds robustness to the approach. This work also raises questions about prior model training contamination, ensuring that the AI/ML community remains vigilant about the integrity of training data in achieving measured outcomes in AI-assisted programming. Overall, this report sets a new standard for evaluating AI model performance in competitive programming environments, paving the way for further exploration in multi-agent systems.