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

Aug 24, 2026

Large language models are widely used for everyday learning, yet typical interactions are unstructured chats rather than curriculum-guided tutoring. Because there is no prior record of the student, the tutor must infer knowledge state from dialogue while simultaneously ordering prerequisite concepts. The work proposes decomposing tutoring into prerequisite graph sequencing plus Socratic exchanges, using a lightweight PPO policy for next-topic selection and dialogue budget. Evaluations across STEM and non-STEM topics show improved mastery rate and fewer dialogue turns than baselines.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages30

Reading Time1 h 16 min

AccessFree to download

Introduction

Self-directed learning in the wild

Prior work on LLM teaching and Socratic tutoring

Core approach: prerequisite knowledge graph and RL-based sequencing

arXiv :2606 . 11744v1 [ cs .CL] 10 Jun 2026 Hey Chat, Can You Teach Me? Structuring Socratic Dialogue for Human Learning in the Wild Sidney Tio∗ School of Computing and Information Systems Singapore Management University [sidney.tio.2021@phdcs.smu.edu.sg](sidney.tio.2021@phdcs.smu.edu.sg) Arunesh Sinha Department of Management Science and Information Systems Rutgers Business School [arunesh.sinha@rutgers.edu](arunesh.sinha@rutgers.edu) Pradeep Varakantham School of Computing and Information Systems Singapore Management University [pradeepv@smu.edu.sg](pradeepv@smu.edu.sg) Abstract Large language models are now widely used for everyday learning, but the underlying interactions are typically unstructured chats rather than following a curriculum. Unlike formal online learning systems, these interactions carry no prior record of the student, so any estimate of what the student already knows must be inferred from the dialogue itself. We show that this gap is not closed by scaling models alone. Frontier and education-tuned LLMs perform poorly when asked to tutor a student over an extended session, because doing so requires three things at once. The tutor must sequence a curriculum, conduct Socratic dialogue, and infer the student’s knowledge state from that dialogue. We propose separating these responsibilities. Given a student query, our system constructs a prerequisite knowledge graph in which subtopics are nodes and dependencies are edges, and frames tutoring as deciding which node to teach next and how many dialogue turns to spend on it before moving on. A lightweight PPO policy handles this sequencing decision, while an LLM conducts the Socratic exchange at the chosen node and returns a signal of student progress. Across held-out STEM and non-STEM topics, our PPO-paired tutor outperforms heuristic baselines, frontier general-purpose models, and a model specialised for Socratic dialogue—on both the rate at which students reach full curriculum mastery and the number of turns required. Explicit curriculum structure delivers gains that scaling the underlying model does not. 1 Introduction Large Language Model (LLM) assistants such as Claude and ChatGPT are increasingly the medium through which people learn outside the classroom. Learning and knowledge-seeking together account for an estimated 30 to 40% of consumer ChatGPT conversations [Chatterji et al., 2025], education is ∗ Corresponding author. Preprint. the second-largest task category on Claude [Handa et al., 2025, Bent et al., 2025], and the share of U.S. teens using ChatGPT for schoolwork doubled from 13% to 26% between 2023 and 2024 [Sidotiet al., 2025] . Unlike a textbook or a search engine, an LLM can hold a conversation, answer follow-up questions, and adapt its explanations dynamically, which makes it a natural tool for self-directed learning and warrants direct examination of how well it actually performs that role. update graph and continue loop until all topics mastered Figure 1: Overview of the RL-based tutoring system. Left: The student query is decomposed into a prerequisite knowledge graph, and is used to keep track of the curriculum. Centre: The RL tutoring policy selects the next topic to teach based on the interaction history with the student. Right: A single turn of the resulting Socratic dialogue, in which the tutor poses a question, the student responds, and the answer is evaluated. We study this in the setting of self-directed learning in the wild, where a user initiates a session with an LLM chat interface and asks a question about something they wish to learn about, with no syllabus, no instructor-designed sequence, and no prior record of the student. The only signal available is the dialogue itself, and a good tutor must use it to do two things at once: deliver content in an order that respects prerequisite relationships between concepts, and figure out where the student currently stands so that content

FAQs

Why do existing LLM tutoring chats underperform in extended learning sessions?

Because they usually lack curriculum sequencing and do not maintain an explicit notion of what the student already knows, which must be inferred from dialogue alone over time.

How does the proposed system model curriculum structure?

It builds a prerequisite knowledge graph from the student query, where subtopics are nodes and dependencies are edges, turning curriculum delivery into graph traversal.

What role does PPO play in the tutoring process?

A lightweight PPO policy decides which next knowledge node to teach and how many dialogue turns to spend there before moving on.

How is learner progress determined during Socratic dialogue?

An LLM conducts the Socratic exchange at the selected node and returns a signal of student progress, guiding the next sequencing decision.

Aug 24, 2026

Upload by Lucas Vance