Hierarchical Skill Graphs for Structured Learning

This direction is split into other directions: cognitive relationship mapping, hierarchical skill dependency, prerequisite-aware recommendation. Its last version stays readable and citable here.

Version 1 of 1276 papers400 conceptsfrozen 2026-10-04

In both human education and machine learning, we often treat knowledge as a flat checklist of items to be mastered. We assume that if a student or a model sees enough examples of "multiplication" and "addition," they will eventually learn both. However, research across decades of artificial intelligence and pedagogy suggests that knowledge is not a collection of isolated silos, but a tower of interdependent building blocks.

The core insight of this research direction is that skills possess a prerequisite structure. If a foundational "building block" is missing, attempting to learn a sophisticated heuristic can actually degrade performance rather than improve it. To solve this, researchers model knowledge as a Directed Acyclic Graph (DAG) or a Skill Graph, where nodes represent discrete concepts and edges define the logical flow of mastery.

The Failure of Independent Skill Modeling

Traditional models in Knowledge Tracing (KT) and Cognitive Diagnosis often treat skills as independent latent variables. As Qi Liu and Zhenya Huang argue in their survey of the field, this "flat" view fails to capture the reality of learning. If a student hasn't mastered a foundational skill, they cannot successfully apply a dependent advanced ability, no matter how much they practice the latter.

This isn't just a problem for human students. Jonathan Schaeffer 0001 and T. Anthony Marsland observed as early as 1985 that adding sophisticated heuristics to an expert system can be counterproductive if the underlying foundational routines are absent. The utility of a specific knowledge module is contingent on the presence of its prerequisites.

Mapping the Curriculum as a Graph

To move beyond flat lists, researchers like Tom Murray 0001 and Beverly Park Woolf pioneered tools to encode knowledge as directed graphs. In these structures, an edge from Concept A to Concept B explicitly states: "You must understand A to make sense of B."

This graphical abstraction allows for several breakthroughs:

  1. Automated Lesson Planning: By traversing the graph, systems can automatically generate pedagogically sound paths. Kwok-Keung Yum and Thomas J. Richards demonstrated that Classification Networks could automate curriculum prescription by indexing the transitions between subjects.
  2. Reachability and Scaffolding: Using a "reachability matrix," as explored by Ana Villanueva and colleagues, instructors can ensure that Augmented Reality instruction is delivered in a sequence that respects foundational dependencies.
  3. Semantic Relatedness: Models like Semantic TrueLearn, developed by Sahan Bulathwela and Emine Yilmaz, use these graphs to solve the "cold-start" problem. If a system knows you are proficient in "Linear Algebra," it can use the graph to predict your proficiency in "Principal Component Analysis" before you've even attempted it.

Applying Skill Hierarchies to Large Language Models

The most recent frontier for this idea is the training of Large Language Models (LLMs). Standard fine-tuning often ignores logical dependencies, throwing all data at the model at once. However, the Skill-it! framework by Jue WANG treats LM training like a human curriculum. By identifying a "natural order" where simpler tasks reduce the data volume needed to master complex ones, Skill-it! achieves higher performance with less data.

Similarly, Jiaao Chen and Diyi Yang introduced Dynamic Skill Adaptation (DSA). Instead of random sampling, DSA uses a skill graph to ensure foundational concepts are introduced before advanced applications. This structured progression allowed models to outperform ChatGPT on specialized benchmarks like Pre-Calculus.

Why It Works: The Efficiency of "Mastering Rate"

The intuition behind the success of these hierarchical models is best captured by the MR algorithm from Lucas Willems and Salem Lahlou. In reinforcement learning, an agent can waste millions of steps trying to solve a "hard" task that is currently impossible because its prerequisites aren't met.

By using a DAG to define prerequisites, the system tracks the "mastering rate" of foundational nodes. It only "unlocks" advanced tasks once the foundation is stable. This prevents the model from getting stuck in "hard-exploration" traps where random chance is the only way to find a solution. As Anna Kravchenko and Rhodri Cusack note, certain high-level abstractions are simply impossible to learn from scratch without the scaffolding provided by prerequisite tasks.

Papers in this version

Showing 60 of 276 papers.

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_hierarchical_skill_graphs_for_structured_57f6f5f7,
  title = {Hierarchical Skill Graphs for Structured Learning},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/hierarchical-skill-graphs-for-structured-learning/rdr_57f6f5f7021c4608b95a0178b0d70e13},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 1, frozen text):

@misc{lacuna_hierarchical_skill_graphs_for_structured_57f6f5f7_v1,
  title = {Hierarchical Skill Graphs for Structured Learning},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/hierarchical-skill-graphs-for-structured-learning/rdr_57f6f5f7021c4608b95a0178b0d70e13/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}