A new research paper reveals how frontier AI models can work together to train smaller coding assistants more effectively than traditional methods.

The study, published on arXiv, introduces a "compete-then-collaborate" framework where four major AI teachers — Anthropic's Claude, OpenAI's Codex-GPT, xAI's Grok, and Google's Gemini — first compete against each other before collaborating to build training curricula.

Researcher Miseong Shawn Kim tested the models using execution-based verification through unit tests and input-output checks, rather than relying on potentially biased AI judges. The target student was Qwen2.5-Coder, available in 7B and 32B parameter versions.

Competition reveals model strengths

When tested on standard coding problems, all four frontier models achieved near-perfect performance after self-correction, scoring 99-100%. However, harder competition problems separated their capabilities significantly.

Gemini led with 77% accuracy, followed by Claude and Codex-GPT tied at 69%, while Grok trailed at 50%. Despite these ranking differences, the student model's final performance remained consistent regardless of which teacher ranked highest.

Imitation learning shows limitations

Traditional supervised fine-tuning on verified solutions from the teacher models actually degraded performance in already-competent students. The 7B model dropped from 76.7% to 72.7% on MBPP-test, while the 32B version fell from 5.9% to 2.9% on competition problems.

This finding challenges conventional wisdom about knowledge distillation from larger to smaller models.

Reinforcement learning delivers gains

Using the same collaborative curriculum in a reinforcement learning environment with verifiable rewards produced opposite results. The student model improved from 5.9% to 8.8% on competition problems — a 49% relative gain.

The research suggests AI teacher collaboration works best when creating verifiable learning environments rather than simply pooling answers for imitation.

The authors released a reproducible pipeline for NVIDIA GB10 systems with framework patches for running GRPO on current AI infrastructure stacks.