> 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/part-iv-synthesis/chapter-5/5-1-synthesis.md). # 5.1 The Four Paradigms in Context The four paradigms of neuro-symbolic AI described in Chapter 4 are not competing alternatives — they are complementary tools for different situations. The key is knowing which to reach for. ## When Symbolic Helps Neural (Section 4.1) This is the **default starting point** for practitioners who already have a working neural system and want to improve its reliability or sample efficiency. Its defining characteristics are: drop-in compatibility (no architectural surgery), safety guarantees via hard constraint layers, and dramatically reduced data requirements when symbolic knowledge is available. **Use it when:** You have domain knowledge that is expressible as logical constraints, your neural system violates these constraints in practice (either during training or inference), and you cannot afford incorrect outputs. Medical imaging, financial modeling, safety-critical control, and any compliance-bound application fall into this category. **Representative systems:** Physics-Informed Neural Networks, Semantic Loss Functions, Neuro-Symbolic Concept Learners. ## When Neural Helps Symbolic (Section 4.2) This is the right pattern when you have a **symbolic system that is correct but computationally intractable**. The neural component serves as a trained approximation of the symbolic system's most expensive subroutines. The key guarantee: incorrect neural guidance means slower performance, never incorrect results — the symbolic oracle verifies everything. **Use it when:** Your symbolic system has strong formal guarantees (a SAT solver, a formal planner, a theorem prover) and the computational bottleneck is a learnable function (heuristic estimation, policy selection, variable ordering). You can generate training data from the symbolic system itself. **Representative systems:** STRIPS-HGN, Neural MCTS (AlphaGo/AlphaProof pattern), NeuroSAT, ∂ILP. ## Hybrid Co-Processing (Section 4.3) This is the **most powerful and most demanding** paradigm — both neural and symbolic components are first-class participants, passing structured representations back and forth in tight loops. It is the right architecture when neither component alone is sufficient: the neural component cannot reason reliably at scale, and the symbolic component cannot process raw perceptual inputs. **Use it when:** Your task requires both raw perception (images, natural language, sensor streams) and formal reasoning (logical inference, constraint satisfaction, verified planning). The interface between components must be a formally specified data structure. Autonomous driving, clinical decision support, program synthesis, and mathematical reasoning all fit this template. **Representative systems:** AlphaProof, GraphRAG, GreaseLM, PLOI, Voyager, ReAct. ### The Fourth Paradigm: Pure Neural World Models (Section 4.4) The fourth paradigm covers approaches without explicit external symbolic structure, split into two distinct sub-paradigms: **§4.4.1 Differentiable Abstract Interpretation (DAI)** — the symbolic abstract domain becomes the neural model's own internal representation space. The Lattice Deduction Transformer (Davis et al., 2026) achieves 100% accuracy on Sudoku-Extreme with 800K parameters (frontier LLMs score 0%), through on-policy training with the alpha operator that approximates sound deduction operators on a formal lattice. **§4.4.2 JEPA World Models** — pure feature-predictive physical planning: V-JEPA self-supervised representations → action-conditioned world model → CEM planning in latent space. No reward, no pixel reconstruction, no formal symbolic structure. V-JEPA 2.1 (Mur-Labadia et al., 2026) achieves +20 percentage points in real-robot grasping over prior baselines. **The shared property:** Both sub-paradigms operate without a separate external symbolic oracle. DAI achieves structure through the abstract lattice architecture itself; JEPA achieves it through hierarchical feature learning. This makes them the natural counterparts to §4.1–4.3. **Use §4.4.1 when:** The problem has a formal lattice constraint structure (CSP, SAT, combinatorial optimization), solution samples are available, and empirical soundness (correct or abstain) is preferred over maximum recall. **Use §4.4.2 when:** Physical planning from perception is required, labeled rewards are unavailable or expensive, and a cost function can encode the task objective more easily than formal preconditions and effects. ## The Convergent Alternative: LeCun's JEPA World Model Program Before the unifying thread, it is worth placing the four paradigms in the context of the most prominent neural-only alternative. Since 2022, Yann LeCun's group at Meta AI has pursued a research program — the Joint Embedding Predictive Architecture (JEPA) — that independently converges on the same diagnosis as this book but proposes a different remedy.(LeCun, 2022) **The convergent diagnosis:** LeCun's "A Path Towards Autonomous Machine Intelligence" (2022) identifies critical limitations of autoregressive generative models: they predict at the wrong level of abstraction, lack a world model for simulating consequences of actions, and cannot perform systematic search over action sequences. His proposed modular architecture has six components — **Configurator** (sets task objectives), **Perception** (encodes sensor inputs), **World Model** (predicts future states), **Short-Term Memory** (maintains world state), **Cost** (evaluates states against objectives), and **Actor** (plans and executes) — and mirrors the logical structure of the neuro-symbolic hierarchy: perception is neural, planning is over a learned forward model, objectives are encoded in a cost function. **The divergent solution:** Where neuro-symbolic AI inserts *formal symbolic representations* (PDDL states, knowledge graph triples, logic programs) as the interface between perception and planning, LeCun's JEPA inserts *abstract learned representations* (latent feature spaces learned by self-supervised feature prediction). By 2026, JEPA-WMs achieve state-of-the-art performance on real robot manipulation (Terver et al., 2026; Mur-Labadia et al., 2026), demonstrating that learned feature spaces can support effective physical planning. **The open question — and why it matters for this book:** Whether JEPA-WMs can satisfy the *formal correctness guarantees* that symbolic planners provide — soundness, decidability, verified safety envelopes — remains unresolved. The stable-worldmodel benchmark (§4.3.5) shows that world model accuracy does not automatically translate to planning reliability. This is precisely the gap that neuro-symbolic verification (§5.2, Problem 3) is designed to fill. Our speculation: the most likely 2029–2031 architecture integrates both — JEPA-style representations for perception and continuous dynamics, formal symbolic structure for goal specification, constraint verification, and safety guarantees — but this convergence remains an open research question. *Reference:* LeCun, Yann. "A Path Towards Autonomous Machine Intelligence." *OpenReview*, Version 0.9.2, June 2022. *** ## The LLM-Modulo Framework as Unifying Thread Across all four paradigms, the **LLM-Modulo** architecture (Section 1.3) serves as a practical unifying principle. The LLM generates proposals; symbolic components verify and constrain. This separation — neural generation, symbolic verification — is not a single system design but a design principle that manifests differently in each paradigm: * In **"Symbolic Helps Neural" (§4.1):** the verifier acts at training time (constraint loss) or inference time (projection layer). The neural component is primary; symbolic knowledge shapes its outputs. * In **"Neural Helps Symbolic" (§4.2):** the neural component generates heuristics that the symbolic system uses but overrides with its own correctness guarantee. Symbolic correctness is never compromised. * In **"Hybrid Co-Processing" (§4.3):** neural and symbolic components alternate roles in a continuous loop, each verifying and improving the other. The interface between them is the critical engineering challenge. * In **"Pure Neural World Models" (§4.4):** the LLM-Modulo principle reaches its most subtle form. In §4.4.1 (DAI), the symbolic structure is embedded *inside* the neural architecture — the lattice is the representation, and the forward pass approximates a sound deduction operator. In §4.4.2 (JEPA), there is no symbolic verifier at all; the system relies on the quality of the learned representation and the cost function. §4.4.2 is the paradigm where departing from the LLM-Modulo principle most visibly trades formal correctness for perceptual generality. The four paradigms form a spectrum from "symbolic dominates" (§4.1 and §4.2) through "bidirectional co-equal" (§4.3) to "no external symbolic component" (§4.4). Chapter 6's decision guide translates this spectrum into actionable selection criteria. *** --- # 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/part-iv-synthesis/chapter-5/5-1-synthesis.md?ask=&goal= ``` `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.