Neurosymbolic architecture: the way out of the parrot-versus-solver compromise

For thirty years, AI research split into two tribes. The connectionists, who view intelligence as an emergent statistical phenomenon, gave us deep learning and LLMs. The symbolists, who view intelligence as the manipulation of formal structures, gave us Prolog, expert systems and ontologies. Each tribe scored its points, and each has its characteristic weaknesses.

Neurosymbolic architecture is the way out of this duel from above: combining the strengths of both paradigms to compensate for their respective weaknesses. It is the most promising direction for building reliable AI agents on specialized domains.

The diagnosis on LLMs alone

An LLM alone produces non-deterministic outputs. Ask the same question twice and you may get two different answers. That's acceptable for creative work, it's disqualifying for auditable decisions.

An LLM alone hallucinates on specialized domains. The further the domain is from generic training corpora, the more quality drops. That's precisely the case for enterprise business logic, which is by definition absent from general corpora.

An LLM alone has no structured memory of the system it manipulates. It doesn't know what it did five minutes ago beyond its context window, and it doesn't know what it will do in five minutes beyond the current prompt. This absence of an internal model is the main obstacle to a reliable agent.

The diagnosis on symbolic systems alone

Conversely, a symbolic system alone is rigid, fragile in the face of ambiguity and imprecision, and expensive to maintain. The expert systems of the 1980s ran into precisely this difficulty: capturing every rule of a domain and keeping them up to date. What works in the lab fails on the diversity of the real world.

The parrot-versus-solver compromise is false. You don't have to choose between an LLM that hallucinates and a formal solver that can't read an email. You can combine the two, and that is precisely what neurosymbolic architecture does.

Kautz's taxonomy, the field reference

Henry Kautz, in his Engelmore Memorial Lecture at AAAI 2022 (published in AI Magazine 43(1)), proposed the taxonomy that has become the field reference. Six architectures are distinguished by the degree and direction of integration between neural and symbolic components. They are not all equivalent for enterprise systems.

Two of the architectures dominate in production in 2026: Neuro | Symbolic, where the neural component converts natural language into a symbolic representation then manipulated by a formal reasoner; and Neuro[Symbolic], where the symbolic engine is embedded in the neural loop, which calls it as a System 2 service. The first is where you find AlphaGeometry at DeepMind and the medical system Plato-3 published in Nature Portfolio in November 2025. The FAOS platform published on arXiv in April 2026 is closer to the second.

AlphaGeometry, the paradigmatic example

AlphaGeometry, published by DeepMind in early 2024, is cited in the literature as the canonical example of successful Neuro | Symbolic coupling. An LLM guides a symbolic deduction engine to solve geometry problems at olympiad level. The neural part generates candidate geometric constructions. The symbolic part verifies each step by formal deduction.

The result is unambiguous: the LLM alone cannot solve the problems, the formal solver alone doesn't know where to start, and the combination reaches expert human level. The value is not in joint learning, it is in the complementarity of the two components.

Plato-3, the precedent in real production

Plato-3, published in Nature Portfolio in November 2025, is closer to a classic enterprise case. The system connects GPT-4 to a Prolog expert system to extract structured clinical parameters from free-text medical reports. GPT-4 proposes candidates, Prolog verifies them against a formal medical ontology. In case of contradiction, Prolog re-prompts GPT-4 with its own reasoning.

The results speak for themselves: F1 = 1.00 for the combined system versus F1 = 0.63 for GPT-4 alone, and 100% accuracy on PSA values. That is the qualitative leap this kind of neurosymbolic coupling produces in a specialized domain.

FAOS and the inverse parametric knowledge effect

The FAOS paper published in April 2026 on arXiv is the most recent reference for a generalist enterprise neurosymbolic system. The platform deploys 650 agents across 21 industry sectors with asymmetric ontological coupling: role, domain and interaction ontologies constrain the agents' inputs.

The FAOS authors empirically establish an important result: the value of symbolic coupling is inversely proportional to the coverage of the domain in the LLM's training data. The more specialized and underrepresented your domain in the corpora, the more value the symbolic ontology brings. That's a strong argument for compliance, niche finance, specialized healthcare, defense.

The four gaps current neurosymbolic systems do not address

Colelough and Regli's review on arXiv in 2025 surveyed 167 papers on neurosymbolic systems between 2020 and 2024. The diagnosis is harsh: production systems remain rare, and most of the work consists of proofs of concept on narrow domains. Four structural gaps stand out.

• Replay determinism: no neurosymbolic production system guarantees that replaying the same pipeline produces the same result. LLMs remain non-deterministic.
• Genericity of the symbolic side: ontologies are built by hand for narrow domains. No system has a generic symbolic component capable of modeling any domain.
• Scalability of the symbolic side: reasoners (Answer Set Programming, Prolog, ILP) saturate on large predicate spaces.
• Multi-agent accountability: when several agents collaborate, traceability of who decided what for what reason is generally absent.