Reproduce “Sherpa: Teaching LLMs to Teach Adaptively” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Weixian Xu, Yanzhe Zhang, Zora Zhiruo Wang, Changyu Chen, Diyi Yang
Sherpa trains a math tutor through multi-turn interactions with simulated students whose instructional preferences affect which guidance they accept, rewarding the tutor by the student’s later problem-solving performance. On the filtered math evaluation, its Qwen3-8B teacher reached 69.0% average post-tutoring accuracy across seven student archetypes, versus 48.6% for the untrained teacher; this does not establish learning gains in real students.
Key insights
- Student learning outcomes can serve as a flexible training signal for teaching, instead of directly optimizing a fixed list of pedagogical behaviors.
- Training across varied instructional preferences improved results even for preference types excluded from training, suggesting transfer to unseen learner needs within the tested setup.
- Gated simulated students make preferences consequential, but the framework’s discrete, enforced archetypes are still an approximation of how real learners behave.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.