Finding the Optimal Granularity of Knowledge Abstractions

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 2161 papers198 conceptsfrozen 2026-10-04

In artificial intelligence, knowledge is rarely useful in its rawest form. A robot that only knows how to "turn a silver-colored 10mm hex bolt at 5 Newton-meters" cannot fix a different machine, while a robot that only knows to "repair things" lacks the instructions to do anything at all. This tension defines the granularity problem: the challenge of finding a level of abstraction that is specific enough to be actionable but general enough to be reusable.

Across domains ranging from mathematical reasoning to biological modeling, researchers are moving away from "flat" data representations. Instead, they are developing methods to navigate hierarchical "conceptual pyramids," seeking a "middle ground" of abstraction that maximizes both logical precision and generalizability.

The Fundamental Tradeoff

The core difficulty in knowledge modeling is a balancing act between two failure modes. Juncai Li and Ru Li describe this as a tradeoff between specificity and utility in their work on Hierarchical Conceptual Entailment Graphs.

  1. Under-abstraction (The Noise Trap): When a system uses hyper-specific categories, it suffers from data sparsity and overfitting. In single-cell RNA sequencing, Jiajie Peng found that using overly specific gene terms leads to fragmented representations that fail to distinguish unique cell functions Combining gene ontology with deep neural networks to enhance the clustering of single cell RNA-Seq data.
  2. Over-abstraction (The Vagueness Trap): Conversely, broad concepts often lack the "discriminative power" needed for reasoning. Lei Zan and Keli Zhang demonstrated in CAMA that if a mathematical causal graph is too coarse, it provides insufficient guidance for solving elite-level competition problems.

Identifying the "Knee Point" of Abstraction

If the extremes are dangerous, how do we find the "Goldilocks" zone? Several researchers have proposed mathematical frameworks to identify the optimal level of detail automatically.

One prominent approach is the Minimum Description Length (MDL) principle. In On Conceptual Labeling of a Bag of Words, Xiangyan Sun and Yanghua Xiao use MDL to summarize sets of words. The goal is to find a label that is concise (short description) but covers the data well (low error). For example, labeling {China, Japan, Korea} as "Asian countries" is more efficient than listing them individually, but more precise than calling them "Geopolitical entities."

Another technique involves the Statistical Knee Point. When mapping metaphors, Mengshi Ge and Erik Cambria applied a knee point algorithm to lexical hierarchies to find concepts that are neither too specific to a single instance nor too broad to be meaningful Explainable Metaphor Identification Inspired by Conceptual Metaphor Theory. This "knee" represents the point of diminishing returns where adding more detail no longer significantly improves the model's explanatory power.

Granularity in Reasoning and Generation

The need for optimal abstraction is particularly acute in Large Language Models (LLMs), which often fall into "reasoning shortcuts." Kaiqiao Han and Yangqiu Song argue that models fail logical challenges like the Winograd Schema because they focus on specific, distracting entities rather than generalized concepts. Their Abstraction-of-Thought (AoT) method forces models to reason at a higher level of the hierarchy to improve robustness Concept-Reversed Winograd Schema Challenge.

This structural guidance also improves translation. Jian Yang found that predicting a "soft template"—a structural skeleton of a sentence—helps neural machine translation, but only if the template is at the right depth of the syntactic tree. Overly detailed trees are too hard to predict, while overly simple ones provide no help Improving Neural Machine Translation with Soft Template Prediction.

The Human-in-the-Loop Perspective

Interestingly, the "optimal" granularity for a machine often aligns with the level at which humans can reach a consensus. In linguistics, Cecily Jill Duffield and colleagues found that WordNet's verb senses were often too fine-grained for humans to agree on. By clustering these into a "middle ground," they improved both human inter-annotator agreement and the accuracy of automatic systems Criteria for the Manual Grouping of Verb Senses.

This suggests that the search for optimal granularity is not just a computational convenience, but a reflection of how knowledge must be structured to be communicable and robust. Whether it is Zhiruo Wang inducing programmatic skills for web agents Inducing Programmatic Skills for Agentic Tasks or Weiyu Liu using attention to navigate type hierarchies Path Ranking with Attention to Type Hierarchies, the goal remains the same: finding the level of abstraction where knowledge becomes a tool rather than a burden.

Papers in this version

Showing 60 of 161 papers.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 136 papers
  2. Version 1 · 2026-10-04 · 161 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_optimal_abstraction_level_59178a50,
  title = {optimal abstraction level},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/optimal-abstraction-level/rdr_59178a50fb6341dd99ec727897bff04e},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_finding_the_optimal_granularity_of_knowl_59178a50_v1,
  title = {Finding the Optimal Granularity of Knowledge Abstractions},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/finding-the-optimal-granularity-of-knowledge-abstractions/rdr_59178a50fb6341dd99ec727897bff04e/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}