Narrow and Deep: An Ontology Tower as the Knowledge of an LLM Agent for an Industrial Equipment System

arXiv:2610.11768 · eess.SY, cs.AI, cs.SY · Submitted 2026-10-08 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.

Rosa: I'm Rosa, and with me are Dev and Taro, guest researcher.

Dev: Today's paper: "Narrow and Deep".

Rosa: The gist The ontology tower, a narrow-and-deep ontology of a single equipment system whose knowledge deepens through physical derivations and through lessons incorporated from the operating journal,

Dev: First, who's behind it and why it matters.

Paper summary: Rosa: So, to recap, this paper proposes the ontology tower as a way to give an LLM agent knowledge about a single industrial equipment system. The central thesis is that this tower deepens its knowledge through two paths: physical derivations from measured points and lessons pulled from the operating journal.

Dev: And they claim this structure is narrow in terms of entities but very deep in what it contains, which they call the ontology tower <ref:2610.11768#pg1>. It’s designed to supply the agent with exactly what it needs for that specific plant environment.

Rosa: They tested this on a real low-humidity air-handling test plant operated by a programmable logic controller, and they showed that when the agent received text projected from this tower, its primary trap avoidance rate went up by about twenty percentage points <ref:2610.11768#pg3>.

Dev: What matters is that they found naming the specific points and giving them their units made the biggest difference, with a change of fifty-four point six percentage points <ref:2610.11768#pg3>. This suggests structure really helps the AI avoid mistakes in this kind of task.

Taro: So from an autonomy standpoint, it’s about moving beyond just big model size and actually grounding the AI's knowledge in concrete, plant-specific details <ref:2610.11768#pg3>. It’s not just about how big the model is, but what specific knowledge it has access to.

Rosa: Right. The paper emphasizes that because plant data is often locked on a private network, building this explicit structure and feeding it to the agent in controlled parts is necessary <ref:2610.11768#pg2>. It addresses the gap of not knowing what knowledge a particular plant should contain and how it should reach the AI.

Dev: And they show that incorporating lessons from operation directly into that ontology, instead of just leaving them in a search tool, makes the agent actually use those learned behaviors <ref:2610.11768#pg3>. That’s a key difference in how the knowledge gets used by the AI.

Rosa: So, for listeners who only listen to this show, it means that if you want an AI to operate industrial equipment reliably, you can’t just rely on its general training; you need to give it a highly specific, layered structure of physical facts and operational history <ref:2610.11768#pg3>.

Dev: And they also showed that this approach allows for live runs where an invariant safety layer checks the agent's proposed actions against known safe lists and ranges, which helped them keep controlled variables in band in twelve out of fourteen runs <ref:2610.11768#pg3>.

Conclusion: Rosa: We’ve talked about how this ontology tower gives agents a structured, deep knowledge base built from physical facts and operational logs <ref:2610.11768#pg1>. The authors, Younghwan Joo and Sung-il Kim, are proposing that this specific knowledge structure is what an agent needs for industrial equipment systems.

Dev: And the implication here is that we’re moving toward giving AI agents context that isn't just abstract text from the internet, but knowledge tightly bound to a physical system <ref:2610.11768#pg2>. It’s about grounding their reasoning in measurable reality.

Rosa: It means for maintenance or operation, we can build systems where the AI doesn't just guess; it understands the couplings and the physics of how things work together in a specific plant <ref:2610.11768#pg3>.

Dev: And they show that by making that knowledge tower evolve with new operational data, you keep the agent current, which is important because plants change over time <ref:2610.11768#pg3>. It’s about maintaining a living knowledge base for the AI.

Taro: From what I see, this isn't just about making models bigger; it’s about providing them with the right kind of structured information tailored to the real-world constraints of equipment operation <ref:2610.11768#pg3>. It’s about control and context.

Rosa: Exactly. So, think of it this way: instead of asking an AI a million general questions, you give it this narrow tower that knows everything relevant to one machine, including the physical rules and what operators have learned from past incidents <ref:2610.11768#pg3>.

Dev: It’s a shift toward supplying context that is explicitly organized around the equipment itself, rather than relying on the agent to piece together everything from its general training <ref:2610.11768#pg2>.

Rosa: That’s the main point of this work, this ontology tower concept, and how it can structure knowledge for agents operating in complex industrial settings <ref:2610.11768#pg3>.

Younghwan Jooa, Sung-il Kima

Korea Institute of Energy Research · University of Science & Technology

eess.SY, cs.AI, cs.SY

Submitted: 2026-10-08

Updated: 2026-10-08

Project page: https://yhjoo-kier.github.io/ontotower

The gist: The gist The ontology tower, a narrow-and-deep ontology of a single equipment system whose knowledge deepens through physical derivations and through lessons incorporated from the operating journal,

Key concepts

Ontology Tower
A narrow-and-deep ontology representing a single equipment system. It deepens in two ways: through physical calculations derived from measured points and by incorporating lessons learned from the operating journal as knowledge nodes. It focuses on being narrow in entities but deep in the known relationships.
Physical Derivations
Knowledge gained by calculating quantities based on measured points using established physical relations. This mechanism allows the ontology to derive new information from raw sensor data, adding a layer of quantitative understanding that goes beyond just reading the initial measurements.
Operating Lessons Incorporation
Integrating specific lessons and experiences from the plant's operating journal directly into the knowledge tower or as records in prompts. This ensures that operational wisdom is actively used by the agent rather than being passively searched, making it more effective for real-world tasks.

Terminology

Summary

The gist The ontology tower, a narrow-and-deep ontology of a single equipment system whose knowledge deepens through physical derivations and through lessons incorporated from the operating journal, was proposed as the knowledge that an LLM agent receives about a plant and it was tested on a real low-humidity air-handling test plant.

How it works

The study proposes an ontology tower, which is described as "a narrow-and-deep ontology of a single equipment system whose knowledge deepens in two ways: through quantities derived from the measured points by physical relations, and through lessons from the operating journal incorporated as knowledge nodes"

This tower is designed to be narrow in entities but deep in what is known

The knowledge deepens through two primary mechanisms:

  1. Through quantities derived from measured points by physical relations

  2. Through lessons from the operating journal incorporated as knowledge nodes

Structure of the Ontology Tower

The ontology tower covers one equipment system, which includes several machines operated as one unit

The structure involves several layers defined by dependencies:

** The bottom stratum P is the equipment schematic **

** The stratum B binds each tag read from the controller to its PLC address, raw scale, unit, and the code that defines it **

** Level 0 holds the tags, the assumptions, and the knowledge nodes **

The height of this tower is computed from declared dependencies; it is neither drawn by hand nor read off the physics

Knowledge Supply and Agent Interaction

The agent does not receive the nodes and edges directly; instead, a text projected from them is placed in the prompt

This projection states in prose and structured blocks, the identity, verification, and known faults of each sensor, the couplings between the units and their automatic logic

The tower can also be supplied to an agent as tools through three readonly queries with which the agent looks up a node, follows its edges, and lists its assumptions

Evaluation Results

In a preregistered evaluation involving nine tasks replayed from plant records using four open-weight models (including 9 to about 750 billion parameters), the projected text raised the primary-trap avoidance rate by about 20 percentage points

The evaluation showed that naming the points and giving their units (K0 to K1) produced the largest change, +54.6 percentage points

The study found that structured knowledge raises the primary-trap avoidance rate

Operating Lessons

Lessons from operation are most effective when incorporated into the tower or placed in the prompt as records rather than left in the journal behind a search tool

When lessons are incorporated, the agent used the lesson

Live Runs

In live runs through an invariant safety layer, agents brought the controlled variable into its target band in 12 of 14 runs

The safety layer checked in order whether the tag is on the list allowed for the run, whether the value lies within the permitted range

Overall, industrial equipment knowledge can be kept as a tower and given to an agent as projected text, and lessons of operation can be kept current by incorporating them into the tower rather than by leaving them for the agent to retrieve

The physical derivations of the tower have yet to be tested with tasks that require them, and extending this approach requires towers produced and updated by extraction

--- Page 24 ---

CRediT authorship contribution statement Younghwan Joo: Conceptualization, Methodology, Software, Validation, Formal analysis, Investigation, Data curation, Visualization, Writing – original draft; Sung-il Kim: Supervision, Project administration Funding acquisition Writing – review & editing.

--- Page 25 ---

Data availability The task specifications the scoring rubric the trial scores and the analysis scripts are available from the corresponding author on reasonable request.

--- Page 26 ---

An earlier public version of the tower of the studied plant, without its operating envelopes, operating risks, and derivation formulas, can be viewed at https://yhjoo-kier.github.io/ontotower/.

--- Page 27 ---

[1] S. Yoon, J. Song, J. Li, Ontology-enabled AI agent-driven intelligent digital twins for building operations and maintenance Journal of Building Engineering 108 (2025) 112802 doi:10.1016/j.jobe.2025.112802

[3] G. Liu, Y. Wang, Y. Lu, H. Wu, K. Liu, L. Liu, J. Chen, Q Fu CoLLM: A retrieval-augmented language model framework for chiller optimization control Journal of Building Engineering 121 (2026) 115683 doi:10.1016/j.jobe.2026.115683

[4] S Lee, M Kang, J Lee, S Lee, D Kim, J Hong, J Shin P Zhang J Ko LLM-based question-answer framework for sensor-driven HVAC system interaction 2025 arXiv:2507.04748 doi:10.48550/arXiv.2507.

Improvements for AI systems

  1. Bold Header: Ontology Tower Knowledge Representation

The system can be supplied with a "narrow-and-deep ontology of a single equipment system whose knowledge deepens in two ways: through quantities derived from the measured points by physical relations, and through lessons from the operating journal incorporated as knowledge nodes." This allows for reasoning that extends beyond raw measurements into derived quantities and operational history.

  1. Bold Header: Knowledge Depth via Physics Derivations

The agent can access quantities derived through engineering relations, such as psychrometric conversions, mass and energy balances, and fan affinity laws, which are held in the tower's levels above measurement, enabling complex system-level calculations without needing the agent to perform those derivations itself.

  1. Bold Header: Knowledge Depth via Operating History Integration

The system can incorporate lessons from operation by writing them into the tower as knowledge nodes linked to the nodes it concerns, ensuring that a lesson incorporated in this way is specific to the system that produced it, and the tower becomes more specific to its system the longer the system is operated.

  1. Bold Header: Knowledge Supply via Projected Text

The agent receives knowledge through a text projected from their tower in a preregistered evaluation of nine tasks, which raised the rate at which the agents avoided the most plausible misjudgment of each task by about 20 percentage points. This structured text provides context that significantly improves performance compared to other supply forms.

  1. Bold Header: Knowledge Supply via Query Tools

The system can use three readonly queries with which the agent looks up a node, follows its edges, and lists its assumptions, allowing it to trace each statement that an agent draws from it to the node it read, providing traceability that text alone cannot offer.

  1. Bold Header: Action Bounding via Invariant Safety Layer

All agent actions are passed through an invariant safety layer which checks constraints such as whether the value lies within the permitted range, whether a person has written to the plant recently, and whether the change satisfies the guard that the plant’s software applies. This ensures agents only attempt setpoint changes that are physically allowed and safe.

Sources

Related papers