Sherpa RL framework trains LLM teachers by directly maximizing simulated student learning outcomes
chrmanning · x · 2026-10-12
The Diyi Yang group's paper "Sherpa: Teaching LLMs to Teach Adaptively" argues that solving a problem isn't the same as teaching it: existing approaches rely on demonstrations, preference data, or predefined pedagogical criteria that aren't grounded in individual student outcomes. Sherpa is a multi-turn RL framework that instantiates multiple LLM-based student archetypes with distinct learning preferences and trains a teacher model to adapt its instruction by directly maximizing their learning outcomes. Teacher LLMs trained with Sherpa improve instructed students' performance across all archetypes, with code released.
More from Research
- Q-CTRL's Boulder Opal automates QPU calibration with 99.95% median 1Q gate fidelity — MJBiercuk · 2026-10-12
- John Wheeler explaining "it from bit": information as the foundation of physical reality — round · 2026-10-12
- Lilian Weng publishes 25-minute deep dive on the math of scaling laws — irinarish · 2026-10-12
- Professor recruiting PhD students at intersection of computational mechanics and computer vision — anand_bhattad · 2026-10-12
- JHU's FracGen teaches video models physics-accurate stretching and tearing from a single image — anand_bhattad · 2026-10-12
- Terence Tao on illustrations in mathematics: beyond proof — smilelamp · 2026-10-12