# Hey Chat, Can You Teach Me? Structuring Socratic Dialogue for Human Learning in the Wild

**Authors:** Sidney Tio, Arunesh Sinha, Pradeep Varakantham  
**arXiv:** [2606.11744](https://arxiv.org/abs/2606.11744)  
**Date:** 2026-06-10  
**Summary page:** https://www.summarizepaper.com/en/paper/2606.11744v1/

## Key points

- Large language models (LLMs) lack a structured curriculum to guide effective learning process
- Scaling up LLMs does not bridge the gap in effective tutoring
- Proposal of a novel approach that separates responsibilities for curriculum sequencing, conducting Socratic dialogue, and inferring student knowledge
- Construction of a prerequisite knowledge graph to guide tutoring process by determining what to teach next and how many dialogue turns to allocate
- Implementation of lightweight Proximal Policy Optimization (PPO) policy for sequencing decisions and leveraging an LLM for Socratic exchanges at each node
- Significant improvements in student learning outcomes across STEM and non-STEM topics achieved by PPO-paired tutor compared to heuristic baselines and specialized models
- Importance of explicit curriculum structure in enhancing learning outcomes beyond model scalability alone, drawing from educational psychology principles.

## Summary

In the study "Hey Chat, Can You Teach Me? Structuring Socratic Dialogue for Human Learning in the Wild," researchers Sidney Tio, Arunesh Sinha, and Pradeep Varakantham delve into the realm of large language models (LLMs) used for everyday learning. They highlight a key issue: while these models engage in unstructured chats with learners, they lack a structured curriculum to guide the learning process effectively. Unlike formal online learning systems that track student progress, LLM interactions rely solely on dialogue to infer a student's knowledge level. The researchers demonstrate that simply scaling up LLMs does not bridge this gap in effective tutoring. To address this challenge, they propose a novel approach that separates the responsibilities of curriculum sequencing, conducting Socratic dialogue, and inferring student knowledge. Their system constructs a prerequisite knowledge graph where subtopics are nodes and dependencies are edges. This framework guides the tutoring process by determining which node to teach next and how many dialogue turns to allocate before moving on. By implementing a lightweight Proximal Policy Optimization (PPO) policy for sequencing decisions and leveraging an LLM for Socratic exchanges at each node, the researchers achieve significant improvements in student learning outcomes across STEM and non-STEM topics. Their PPO-paired tutor outperforms heuristic baselines and specialized models for Socratic dialogue by enhancing both the rate at which students master the curriculum and reducing the number of turns required. Drawing from educational psychology principles emphasizing reasoning through intermediate steps, strategic content sequencing, and active student engagement under uncertainty for improved retention, this study underscores the importance of explicit curriculum structure in enhancing learning outcomes beyond model scalability alone.