LLM-Guided Reinforcement Learning for Mastery Learning: Large-Scale Field Evidence
Talkat Bridging Prediction and Intervention Problems in Social Systems (Simons Institute, 2026) · 2026-01-12..01-18·
summary
A city-scale RCT enrolls ~1,000 Taipei high schoolers where everyone uses a GPT‑4 LLM-powered AI tutor with pedagogical guardrails and the randomization compares a fixed problem sequence to a POMDP-based adaptive sequence that uses LLM-extracted features (meaningful code edits, chat quality, attempts, time) as observations. The platform draws from a TA-verified 40-problem bank per module, GPT‑generated materials, and yields a 0.15 SD improvement for the personalized arm, driven by more time, attempts, and larger gains for beginners and lower-tier schools.
AI-written summary · about
chapters (14)
- 0:00Introduction
- 1:26Speaker background
- 2:53AI skill atrophy
- 7:01Mastery learning in AI
- 9:50Research motivation
- 14:46Study timeline
- 16:34Platform data
- 18:45POMDP fitting
- 19:27Adaptive sequencing
- 22:26Randomized controlled trial design details
- 25:27Treatment vs control
- 27:38Main results
- 30:51Results interpretation
- 31:56Q&A
related talks
25mWhy LLMs Still Fail: Reasoning, Robustness, and Safety
19mLLMs reasoning and agentic capabilities over time
19mRL for Autonomous Coding
27mBuilding Mini Minds Our Adventures in Teaching Small LLMs to Reason
25mDesigning and Building Custom Reinforcement Learning Environments for Fine-tuning LLMs
42mQuestions for Theory in the New Age of Machine Learning
46mPlaying the long game
12mKeynote: How LLMs Have Changed the Way Organizations Do AI