AI Teachers Collaborate to Build Verifiable Curriculum for Coding Students

Current multi-teacher knowledge distillation methods often merge outputs without truly assessing which frontier model provides the best teaching, sometimes relying on biased LLM judges. A new "compete-then-collaborate" framework addresses this by first ranking four frontier AI teachers (Claude, Codex-GPT, Grok, Gemini) using an unbiased, execution-based judge that employs unit tests and stdin-stdout checks. After ranking, these teachers then collaborate to construct a verifiable curriculum specifically for a coding student model (Qwen2.5-Coder). The study yielded three key findings. Firstly, while all teachers achieved near-perfect scores on standard problems after self-correction, harder competition problems revealed performance differences, with Gemini outperforming others. However, the student's robust results were independent of this initial teacher ranking. Secondly, simple imitation learning (SFT) on verified solutions did not improve, and sometimes even degraded, an already competent student model. Crucially, the third finding demonstrated that using the same collaborative curriculum within a reinforcement learning with verifiable rewards (RLVR) environment significantly improved the student's performance on competition problems, achieving a 49% relative gain. This indicates that the true value of AI teacher collaboration lies not in merely pooling answers for imitation, but in jointly creating a verifiable environment where the student actively learns by doing. The researchers also released a reproducible on-prem pipeline for this work.