Why LLMs Fail at Critical Decisions: Ontologies Provide the Missing World Model

LLMs Talk Well but Advisors Don't Trust

After 2023, large language models entered the military-industrial circle. The author's team integrated an LLM into an underwater simulation system where multiple underwater robots perform patrol, search, and engagement tasks. Commanders would ask: "If an enemy submarine maneuvers at this position, what is our optimal response strategy?" The LLM's answers were professionally phrased and logically structured, but operational staff immediately recognized: this result cannot be trusted. The problem was not intelligence — the model was too good at producing plausible-sounding answers. Its understanding of the battlefield world was assembled statistically, not built through formal modeling.

Four Mainstream Approaches Tried and Falsified

To solve this, the author tested four mainstream approaches:

Prompt engineering: Hundreds of prompts injected doctrines, principles, and parameters into context. Changing the phrasing of the same question changed the conclusion. Prompts patch behavior; they cannot guarantee correct reasoning by constraining output.

Agents: Decomposed into reconnaissance, decision, evaluation agents. Individual agents worked, but command relationships, coordination constraints, and priority conflicts among battlefield entities were not "understood" — agents merely passed messages.

RAG (Retrieval-Augmented Generation): Vectorized doctrines, cases, manuals. Retrieved content was always "relevant but insufficient." Battlefield decisions require precise semantics and causal reasoning; RAG does similarity matching — it returns "documents most like an answer," not "the correct answer."

Knowledge graphs: Built entities, relations, triples carefully. Static graphs are "dead": they record "what exists" but do not prescribe "what must be" or "what is allowed." Without axioms, rules, or temporal logic, they cannot derive the new knowledge battlefield decisions need.

All four directions hit a dead end.

The Answer: LLMs Lack a World Model

The breakthrough came during the project's toughest phase. Human experts judge accurately because they possess structured knowledge: hierarchical relationships among combat concepts, equipment performance boundaries, causal chains of action sequences, physically feasible combinations, and doctrine-compliant actions. This is called ontology — explicit, formalized, reasoning-capable modeling of domain knowledge. LLMs provide "statistical intuition"; ontologies provide "structured cognition." The former makes AI speak; the latter makes AI speak correctly. Serious decision scenarios require both.

OntoL: An Ontology-Driven AI Decision Modeling Product

In February 2026, the author began systematic ontology engineering research; by May, OntoL v1 was released. It is an independently developed product, not owned by any organization. Its technical roots are the complete trial-and-error record from years of underwater robot battlefield AI decision projects at CAS. OntoL does one thing: enable LLM reasoning on top of ontologies.

A Real Underwater Battlefield Case

Scenario: Red-force multiple underwater robots conduct barrier patrol in a strait. One robot's passive sonar detects a suspicious contact, suspected enemy submarine.

LLM's first reaction: "Recommend all robots high-speed maneuver to contact point for encirclement." Sounds aggressive, concentrates force — typical "plausible nonsense."

Ontology intercepts step by step:

Barrier continuity rule: Barrier line must not have coverage gaps. All robots leaving to pursue creates a huge gap; target can escape through it. LLM lacks the concept "barrier must be continuous."

Stealth rule: High-speed maneuver raises self-noise, degrades own detection, effectively telling opponent "I'm here." Doctrine: patrol speed must not exceed detection performance inflection point.

Contact procedure: Doctrine requires suspected contact to first close-in identify, reach confidence threshold, then transition to attack. LLM skipped identification, directly suggested encirclement.

Ontology-constrained plan: Nearest robot low-speed advances to identify; others maintain barrier. After identification confidence meets threshold, decide per contact procedure. Every recommendation carries its provenance — which rule, which concept definition. Commanders scan the evidence chain and know why the plan is trustworthy. This is ontology's value: not making AI more articulate, but giving AI's words boundaries and evidence.

How Ontologies Are Generated: A Six-Step Process

Many ask: does ontology require expert teams writing for years? The author's actual workflow, six steps:

Collect knowledge sources: Combat doctrines, equipment manuals, typical cases, expert interview records, historical simulation debriefs — all obtainable text materials.

Concept extraction: Identify core concepts: platforms (underwater robots, submarines, surface ships), sensors, tactical actions (search, track, identify, attack), environmental elements (thermocline, seabed terrain, sound velocity profiles), targets.

Define relations and attributes: Classification hierarchies, mounting relations, performance boundaries. E.g., "certain sensor affected by thermocline," "certain action depends on stealth state."

Rule formalization (most critical): Translate every "must, prohibit, allow" in doctrines into machine-decidable rules. "Non-essential active detection not used" becomes a precondition rule for active detection use.

Simulation validation: Replay historical scenarios to test ontology: system rulings vs expert rulings, count false positives/negatives, expert review each item.

Iterative feedback: Every issue found in simulation feeds back to fix ontology. Rule conflicts, missing concepts, boundary errors — fix, then retest.

Ontologies are not built once; they are "fed" by repeated simulations, becoming more accurate with use. OntoL supports this entire chain: modeling, rule management, consistency checking, LLM collaborative reasoning, conclusion traceability.

Battlefield Pits Are the Same as Industry Pits

People ask: what does underwater robot battlefield decision have to do with enterprise? The isomorphism lies in the problem structure. If your industry has:

Complex business rules that must be obeyed;

Long decision chains, many entity constraints;

High cost of error, "sounds right" is far from enough;

Valuable expert experience that is hard to capture and transfer —

then the pits you hit with RAG and pure LLM solutions are the same batch the author hit on the battlefield. Ontology-driven AI is not about making AI flashier; it's about making AI trustworthy for the first time.

Closing Thoughts

The author doesn't stack demo features or promise "plug-and-play intelligence." He believes a simple principle: first build knowledge right, then let AI reason. If your team struggles with LLM hallucinations, inconsistency, and explainability — let's talk. The pits he stepped in may save you three years.

About OntoL: An independently developed ontology-driven AI decision modeling product, originating from long-term engineering practice in CAS underwater robot battlefield AI decision and underwater simulation deduction domains.