> For the complete documentation index, see [llms.txt](https://neurosymbolicai.gitbook.io/neuro-symbolic-ai-in-practice/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://neurosymbolicai.gitbook.io/neuro-symbolic-ai-in-practice/readme.md).

# Preface

**Community:** [Neuro-Symbolic AI Communities](https://linktr.ee/symbolicai)

**Revision:** 0.7 · June 2026

**Author:** [Vitaly Mirkis](https://www.linkedin.com/in/vitalymirkis/) · VP AI · TrueLogic.ai

***

> *"The test of a first-rate intelligence is the ability to hold two opposed ideas in mind at the same time and still retain the ability to function."*
>
> — F. Scott Fitzgerald, *The Crack-Up*, 1936

***

## About This Book

Artificial intelligence is at an inflection point. Large language models write code, generate images, and converse fluently across virtually every domain. Yet beneath this surface of impressiveness lies a stubborn gap — the gap between *language about the world* and *reliable reasoning within it*.

**Neuro-Symbolic AI** bridges that gap. It combines the pattern-recognition power of modern neural networks with the formal guarantees of symbolic reasoning (e.g., AI Planning, Lean, CSP, SAT, etc.), producing systems that are simultaneously *expressive* and *correct* — systems that understand unstructured natural language and still produce plans that a formal verifier will certify.

**Neuro-Symbolic AI in Practice** e-book is for practitioners who have shipped machine-learning systems and noticed the gap. It is for researchers who want a precise, citable reference at the frontier of the field. It is for teams building autonomous agents that cannot afford to hallucinate action sequences, violate physical constraints, or simply guess wrong when the cost of being wrong is high.

***

## Navigate the Book

* [→ Table of Contents](/neuro-symbolic-ai-in-practice/table-of-contents.md) — All parts, chapters, and sections at a glance
* [→ Start Reading — Chapter 1](/neuro-symbolic-ai-in-practice/part-i-motivation/chapter-1.md) — Why pure neural approaches are insufficient
* [→ Decision Guide — Part V](/neuro-symbolic-ai-in-practice/part-v-decision-guide/chapter-6-decision-guide.md) — Which architecture is right for your problem?
* [→ Glossary](/neuro-symbolic-ai-in-practice/back-matter/glossary.md) — Key terms and definitions

***

## How to Read This Book

The book is organized in five parts across six chapters. Read Parts I–III in sequence for a complete grounding; practitioners who need a specific architecture can jump to Chapter 4 after reading §1.4 (the comparison table). Chapter 6 (the Decision Guide) is designed to be consulted independently at any time.

***

**Part I — Motivation** (Chapter 1) makes the case for why pure neural approaches are insufficient and why the time for neuro-symbolic integration is now. The argument is empirical and architectural, not ideological.

Chapter 1 documents the gap on four fronts: planning benchmarks (PlanBench, TravelPlanner at 0.6% success rate for GPT-4-Turbo), compositional generalization (SCAN, COGS), abstract reasoning (ARC-AGI, with the scale inversion result showing an 800K-parameter LDT beats frontier LLMs that score 0%), and the structural blind spot of probability-based verification (smooth falsehoods). Notably, the chapter shows that even Yann LeCun — VP and Chief AI Scientist at Meta AI, one of deep learning's founders — independently arrives at the same diagnosis from within the neural-primary tradition: LLMs lack world models, cannot plan, and use the wrong architecture.

The LLM-Modulo framework (Kambhampati et al., 2024) provides the constructive alternative: LLMs as collaborative contributors to a verified planning loop, not autonomous reasoners. Extended thinking (o1/o3) is analyzed as an important development that improves the *neural* side of this loop without eliminating the need for symbolic verification.

***

**Part II — Background** (Chapters 2–3) provides the historical grounding and formal foundations that every practitioner needs.

Chapter 2 surveys the full arc from classical AI planning successes to 2026 frontier deployments. The classical milestones — Remote Agent (NASA Deep Space 1, 1999), SPIKE (Hubble Space Telescope scheduler, 1993–present), MAPGEN (Mars Exploration Rovers), CASPER (Earth Observing One) — demonstrate that formal planning systems are production-proven in the hardest environments humans have built. Modern neuro-symbolic milestones include LLM+P, SayCan (CoRL 2022), Code as Policies, RT-2, AlphaCode, AlphaProof (silver medal at IMO 2024), and HAIMEDA (first production neuro-symbolic LLM verification system in regulated medical AI, 2026). The chapter also covers JEPA Physical Planning (Meta AI, 2024–2026) as the neural-only alternative converging on the same physical planning challenge. Agent patterns — ReAct, Reflexion, Tree of Thoughts, Toolformer, Voyager, LATS — are analyzed as practical deployment patterns with their neuro-symbolic readings.

Chapter 3 establishes the formal foundations: the AI planning problem (STRIPS/PDDL), computational complexity (PSPACE-completeness of propositional planning, NP-hardness of oversubscription planning), HTN planning (EXPTIME-completeness of total-order HTN, per Alford et al. 2016), planning extensions (temporal, numeric, conformant, multi-agent, probabilistic), and the connections between HTN decomposition and the Options framework from hierarchical reinforcement learning.

***

**Part III — Core Approaches** (Chapter 4) is the intellectual core of the book. It covers four architectural families with worked examples, implementation pointers, and code references.

**§4.1 Symbolic Helps Neural** — symbolic knowledge constrains, guides, or structures neural learning: Semantic Loss and projection layers, Physics-Informed Neural Networks (PINNs), AlphaFold 2 (SE(3)-equivariant computation via IPA achieving +20 CASP rank points) and RFdiffusion (de novo protein design with SE(3)-equivariant symbolic priors, Watson et al., 2023), SPO+ and CombOptNet for combinatorial optimization, EIDOKU (CSP-based structural hallucination verification), **Concept Bottleneck Models** (CBMs, Koh et al., 2020 — the dominant NeSy approach for interpretable medical AI, with symbolic concept bottleneck and expert intervention), **Neural Module Networks** (NMNs, Andreas et al., 2016 — compositional symbolic programs for visual reasoning, predecessor to NS-CL), the Neuro-Symbolic Concept Learner (NS-CL on CLEVR), Abductive Learning, RLHF/Constitutional AI, **Control Barrier Functions / STL monitors** (CBF, Ames et al., 2019 — formal safety guarantees for neural controllers, deployed in autonomous vehicles and robotics), and **REINVENT** (Blaschke et al., 2020 — LLM-generated drug candidates constrained by multi-property symbolic scoring functions, deployed at AstraZeneca).

**§4.2 Neural Helps Symbolic** — neural networks accelerate, approximate, or replace computationally expensive symbolic components:

* *§4.2.1 Differentiable Reasoning:* Markov Logic Networks, Logic Tensor Networks (LTNs with product t-norm and p-mean quantifiers), End-to-End Differentiable Proving, DeepProbLog, Scallop, ∂ILP, FlashFill, DreamCoder, Popper, ILASP; and **PAL / Program of Thoughts / VisProg / ViperGPT** (Gao et al., 2023; Gupta & Kembhavi, 2023 — the most practically deployed NeSy pattern: LLM generates program, symbolic interpreter executes and returns ground-truth result)
* *§4.2.2 Neural Subroutines:* STRIPS-HGN (hypergraph networks for planning heuristics), AlphaZero-inspired MCTS (PUCT selection), HyperTree Proof Search (HTPS), Tactician, LATS, Neural Combinatorial Optimization (Pointer Networks, Attention Model, GFlowNets), NeuroSAT, SATNet; **AI Feynman / PySR** (Udrescu & Tegmark, 2020; Cranmer, 2023 — neural-guided symbolic regression for scientific equation discovery); **Autoformalization / Llemma** (Azerbayev et al., 2024 — LLM translating informal mathematics to Lean 4 for formal verification)
* *§4.2.3 Answer Set Programming:* NeurASP, DeepStochLog, SLASH

**§4.3 Hybrid / Co-Processing Architectures** — neural and symbolic components as co-equals, exchanging structured information in tight loops: AlphaProof (AlphaZero-inspired MCTS-based RL with Gemini LM + Lean 4 type checker), AlphaGeometry 1 and 2 (DDAR + language model), Knowledge Graph integration (TransE, RotatE, R-GCN, ERNIE, GreaseLM, GraphRAG, QA-GNN, DRAGON, ConceptNet, ATOMIC, COMET), AlphaCode, PLOI, FunSearch, PDDLStream / TAMP, **DreamerV3** (latent pixel-space world model frequently paired with symbolic task planners), **Eureka** (Ma et al., 2024 — LLM iteratively designs symbolic reward functions, verified by RL training oracle, surpassing human-engineered rewards on 83% of dexterous manipulation tasks), **SWE-bench / SWE-agent** (Jimenez et al., 2024 — test suites as formal correctness oracles for software engineering; 50%+ issue resolution with oracle vs. \~2% without), **Selection-Inference** (Creswell et al., 2023 — formally auditable alternating select/infer reasoning chains), and **ProgPrompt** (Singh et al., 2023 — task plans as Python programs with assertion-based symbolic feedback).

**§4.4 Pure Neural World Models** — approaches without explicit external symbolic structure, split into two sub-paradigms:

* *§4.4.1 Differentiable Abstract Interpretation (DAI):* The symbolic abstract domain is embedded as the neural representation itself. The Lattice Deduction Transformer (LDT, Davis et al., 2026) achieves 100% accuracy on Sudoku-Extreme with 800K parameters while frontier LLMs score 0%, remaining empirically sound (correct or abstain). Covers abstract interpretation theory (Cousot & Cousot, 1977), LDT architecture, historical lineage from SATNet, and a design guide.
* *§4.4.2 JEPA World Models:* Feature-predictive physical planning — V-JEPA self-supervised representations → action-conditioned world model → CEM planning in latent space, without reward or pixel reconstruction. V-JEPA 2.1 (Mur-Labadia et al., 2026) achieves +20pp real-robot grasping. Engineering design guide from Terver et al. (TMLR 2026).

***

**Part IV — Synthesis** (Chapter 5) draws together the four paradigms and situates them relative to the broader AI landscape.

§5.1 introduces LeCun's JEPA program as the "convergent alternative" — the neural-only approach that independently converges on the same diagnosis (LLMs cannot plan) but proposes a different remedy (learned world models in abstract feature space vs. formal symbolic structure). The open question — whether learned world models can satisfy correctness guarantees without formal verification — is framed as the defining research challenge at the frontier of both communities.

§5.2 surveys seven open research problems: (1) Reliable Grounding, (2) Knowledge Acquisition at Scale, (3) Scalable Verification (Marabou, α,β-CROWN), (4) Distribution Shift in Learned Components, (5) Formal Semantics for LLM Outputs (smooth falsehoods, EIDOKU, HAIMEDA), (6) Grounding Causal Reasoning in Formal Structure, and (7) Privacy Violations as a Distinct Failure Mode in Data-Sensitive Deployments.

§5.3 provides a five-year outlook (2026–2031), tracking three major trajectories: JEPA physical planning maturing from laboratory to production; neuro-symbolic verification becoming viable for safety-critical domains; and the long-term convergence of continuous world models with formal symbolic structure.

§5.4 is a practitioner's starting guide with concrete tool recommendations, workflow templates, and code links for all major paradigms.

***

**Part V — Decision Guide** (Chapter 6) is a standalone reference tool answering the single most practical question: *which neuro-symbolic architecture is right for my problem?*

A four-step decision table leads to one of six recommendations (Pure Neural, Lightweight Structured, §4.1, §4.2, §4.3, §4.4.1 DAI, or §4.4.2 JEPA-WMs), each grounded in concrete trigger conditions. A domain-by-domain recommendation matrix covers robotics, theorem proving, medical diagnosis, logistics, physics simulation, competitive programming, knowledge-intensive QA, autonomous driving, drug discovery, game playing, HTN task planning, SAT/constraint solving, and lattice-structured CSP. Warning signs (when not to use neuro-symbolic AI) and a pre-commitment audit checklist complete the chapter.

***

## Who This Book Is For

**Practitioners** who have deployed ML systems and are confronting the reliability gap: systems that are fluent but not sound, fast but not correct, capable but not trustworthy. You need formal guarantees and you need a path to get there without abandoning everything that works.

**Researchers** entering the neuro-symbolic field who need a comprehensive, citable reference across the entire landscape — from STRIPS complexity through DreamerV3/JEPA-WMs through the Lattice Deduction Transformer, with every major claim sourced to primary literature.

**Teams** building next-generation autonomous agents — in robotics, scientific discovery, enterprise automation, or any domain where failure has irreversible consequences.

**Prerequisites:** Comfort with machine learning fundamentals (gradient descent, transformers, embeddings) and at least passing familiarity with logic or formal methods. No prior knowledge of AI Planning or knowledge graphs is required — we build both from the ground up, with rigorous definitions, worked examples, and exercises at each level.

***

## Citation and Scholarship Standards

All major claims are backed by citations. References follow **MLA format**, augmented with links to preprints, code repositories, datasets, and model hubs. Citations appear inline as author-year parentheticals `(Author et al., Year)` and are collected in chapter-level `## References` sections. Every inline `*Reference:*` block contains the complete MLA entry.

Highly cited, peer-reviewed primary sources are preferred. Where results are from arXiv preprints, this is noted explicitly. Benchmarks, empirical claims, and performance numbers are attributed to the specific conditions under which they were obtained.

***

*For questions, corrections, or contributions, contact the author via the community link:*

**Community:** [Neuro-Symbolic AI Communities](https://linktr.ee/symbolicai)


---

# Agent Instructions
This documentation is published with GitBook. GitBook is the documentation platform designed so that both humans and AI agents can read, navigate, and reason over technical content effectively. Learn more at gitbook.com.

## Querying This Documentation
If you need additional information that is not directly available in this page, you can query the documentation dynamically by asking a question.

Perform an HTTP GET request on the current page URL with the `ask` query parameter, and the optional `goal` query parameter:

```
GET https://neurosymbolicai.gitbook.io/neuro-symbolic-ai-in-practice/readme.md?ask=<question>&goal=<endgoal>
```

`ask` is the immediate question: it should be specific, self-contained, and written in natural language.
`goal` is optional and describes the broader end goal you are ultimately trying to accomplish on behalf of the user. GitBook uses it to tailor the answer towards what is most useful for that goal.

The response will contain a direct answer to the question and relevant excerpts and sources from the documentation.

Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.