Back to AI Research

AI Research

Compete Then Collaborate: Frontier AI Teachers Buil... | AI Research

Key Takeaways

  • Compete Then Collaborate: Frontier AI Teachers Build a Verifiable Curriculum to Improve a Coding Student Beyond Imitation This research explores how to effec...
  • Large language models increasingly serve as teachers generating training data for smaller students.
  • Prior multi-teacher knowledge distillation methods merge outputs without determining which frontier model teaches best, often relying on an LLM judge biased toward its own outputs.
  • (2) Imitation (SFT) on verified solutions does not improve, and can degrade, an already-competent student at 7B and 32B (e.g., from 76.7% to 72.7% on MBPP-test, and 5.9% to 2.9% on competition problems).
  • The value of AI-teacher collaboration lies not in pooling answers to imitate, but in jointly constructing a verifiable environment where the student learns by doing.
Paper AbstractExpand

Large language models increasingly serve as teachers generating training data for smaller students. Prior multi-teacher knowledge distillation methods merge outputs without determining which frontier model teaches best, often relying on an LLM judge biased toward its own outputs. We introduce a compete-then-collaborate framework where four frontier AI teachers (Claude, Codex-GPT, Grok, Gemini) are ranked head-to-head by an execution-based judge (unit tests and stdin-stdout checks) with fairness controls, and then collaborate to build a verifiable curriculum for a student (Qwen2.5-Coder). We report three findings. (1) Under execution verification, all teachers solve standard problems near-perfectly after self-correction (99-100%) due to a saturation effect, but harder competition problems separate them (Gemini 77% > Claude 69% = Codex 69% > Grok 50%); however, the robust student-side results do not depend on teacher ranking. (2) Imitation (SFT) on verified solutions does not improve, and can degrade, an already-competent student at 7B and 32B (e.g., from 76.7% to 72.7% on MBPP-test, and 5.9% to 2.9% on competition problems). (3) Using the same collaborative curriculum as a reinforcement learning with verifiable rewards (RLVR) environment improves the student (from 5.9% to 8.8% peak on competition problems, a +49% relative gain), reversing SFT's direction. The value of AI-teacher collaboration lies not in pooling answers to imitate, but in jointly constructing a verifiable environment where the student learns by doing. We release a reproducible on-prem pipeline (NVIDIA GB10) with framework patches for running GRPO on a bleeding-edge stack.

Compete Then Collaborate: Frontier AI Teachers Build a Verifiable Curriculum to Improve a Coding Student Beyond Imitation
This research explores how to effectively use powerful "frontier" AI models—such as Claude, Codex, Grok, and Gemini—to train smaller, more efficient coding students. Rather than simply having a student copy the answers provided by these teachers, the authors propose a "compete-then-collaborate" framework. This approach shifts the focus from imitation to creating a verifiable environment where the student learns by solving problems and receiving objective feedback, ultimately improving the student's coding performance.

Ranking Teachers Through Execution

To determine which models are the most effective teachers, the researchers moved away from using other AI models as judges, which can be biased toward their own outputs. Instead, they used an execution-based judge: the teachers were given coding tasks, and their solutions were tested against hidden unit tests and input-output checks. While all models performed near-perfectly on standard, easier problems due to benchmark saturation, the harder competition-level problems revealed clear differences in performance. Gemini led the group, followed by Claude and Codex, with Grok trailing. Importantly, the researchers found that the student's ultimate success did not depend on which specific teacher was ranked highest.

Why Imitation Fails

The study compared two ways of using the teachers' verified solutions. In the first method, the student model (Qwen2.5-Coder) was trained using Supervised Fine-Tuning (SFT) to imitate the teachers' code. The results were counterintuitive: imitation actually degraded the performance of the student models, both at the 7B and 32B parameter scales. The researchers suggest that because these student models are already competent, simply mimicking the teachers' style does not provide the necessary growth and may even introduce knowledge conflicts.

Learning by Doing with RLVR

The second method used the same set of verified solutions to create a Reinforcement Learning with Verifiable Rewards (RLVR) environment. In this setup, the student model is not asked to copy the teacher's code; instead, it is rewarded when its own code passes the execution tests. This approach successfully reversed the negative trend seen in imitation. By treating the verified problems as a landscape for policy optimization, the student model showed a significant performance gain—up to a 49% relative improvement on held-out competition problems—proving that the value of AI collaboration lies in building an environment where the student learns through practice.

Key Considerations

The researchers emphasize that their findings are specific to the Python coding tasks and the Qwen2.5-Coder model family used in the study. They also note that while the RLVR gains are clear, the absolute improvements are modest, and the performance can fluctuate during training, suggesting that careful checkpoint selection is important. Finally, the authors highlight that their work is strictly academic and designed for reproducibility, providing a full on-prem pipeline and framework patches to allow others to verify their results or conduct similar experiments.

Comments (0)

No comments yet

Be the first to share your thoughts!