Improving Artificial Teachers by Considering How People Learn and Forget
summary
A model-based tutoring system treats teaching as a POMDP, uses a simple exponential forgetting model with two parameters inferred by Bayesian online updates, and runs planners that account for a learner’s practice schedule. Tested on simulated agents and N = 53 humans learning Japanese kanji, the myopic planner increases items recalled but presents far more items so the learned/seen ratio falls, while a conservative planner matches the Leitner baseline. Implementation shows modular online inference and planning with concrete tradeoffs in workload versus final recall.
AI-written summary · about
related talks
11mThe Unreasonable Effectiveness of Prompt Learning
08mReal AI Agents Need Planning, Not Just Prompting
02h 02mTeaching Our AI to ... Learn
14mGenAI Superpowers EdTech
51mMemory Masterclass: Make Your AI Agents Remember What They Do!
14mUnconference: Self-optimizing deep research agents
26mGrappling with the Bitter Lesson
16mUser Interaction Models for Disambiguation in Programming by Example