> 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-iii-core-approaches/chapter-4/4-4-pure-neural/4-4-differentiable-abstract-interpretation/4-4-design-principles.md).

# 4.4.1.3 The Scale Inversion Result

The comparison — 800K parameters achieving 100% vs. frontier LLMs at 0% on the same benchmarks — is the sharpest known empirical demonstration of the book's central thesis and deserves careful analysis.

**Why frontier LLMs score 0%.** Sudoku-Extreme instances require genuine multi-step deduction: no statistical correlation in any training dataset maps reliably to the answer. LLMs trained on text containing solved Sudoku puzzles learn surface-level correlations, not the underlying constraint propagation algorithm. Scaling up the model does not add this capability — it scales pattern recognition, not systematic search.

**Why LDT achieves 100%.** LDT's lattice structure prevents the model from producing any output state that violates the abstract domain's invariants. Each forward pass must produce a *more informative* (lower in the lattice) abstract state — eliminating candidates that are demonstrably inconsistent. The architecture enforces the type discipline that deductive reasoning requires. The transformer cannot "hallucinate" a solution that violates constraints because the lattice encoding makes constraint-violating states unrepresentable.

**The architecture primacy principle.** For structured combinatorial reasoning, architectural choices dominate model scale by orders of magnitude — LDT's 800K parameters achieves what systems with tens to hundreds of billions of parameters cannot. This does not mean LLMs are useless — it means they are the wrong tool for tasks that require systematic, sound deduction over structured state spaces. The decision guide in Chapter 6 (§6, Step 1) is precisely calibrated to this distinction: if provably correct outputs are required and the problem has a formal constraint structure, §4.4.1 (or §4.2) is the right paradigm.

***

> **When to Use §4.4.1 (Differentiable Abstract Interpretation)**
>
> Use DAI when all three of the following hold:
>
> 1. **The problem has a formal lattice constraint structure** — a constraint satisfaction problem, SAT instance, combinatorial puzzle, or other domain where candidate sets per variable can be partially ordered.
> 2. **Solution verification is available** — you can check whether a proposed complete assignment is valid (a Sudoku checker, a SAT oracle, a type checker).
> 3. **Empirical soundness is more important than maximum recall** — you prefer a system that returns a correct answer or abstains over one that always outputs something but may be wrong.
>
> **Do not use DAI when:** the problem lacks formal lattice structure (continuous optimization, natural language generation, image classification); you need guaranteed completeness rather than soundness; or the constraint structure is so complex that designing the abstract domain itself becomes the harder engineering problem than training a system directly.

***


---

# 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-iii-core-approaches/chapter-4/4-4-pure-neural/4-4-differentiable-abstract-interpretation/4-4-design-principles.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.