> 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.md).

# 4.4.1 Differentiable Abstract Interpretation

> *"A fourth architectural paradigm: the symbolic domain IS the neural representation."*

***

The three paradigms in §4.1–4.3 share a structural assumption: neural and symbolic components are *distinct modules* that interact through a defined interface. **Differentiable Abstract Interpretation (DAI)** dissolves this distinction entirely — the symbolic abstract domain of a constraint solver becomes the neural model's own internal representation space.

> **Practitioner note:** DAI is the most technically demanding of the four paradigms. It is the right choice when the problem has a formal lattice constraint structure (CSP, SAT, combinatorial optimization), solution verification is available, and empirical soundness (correct or abstain) is preferred over maximum recall. See the "When to Use" callout at the end of §4.4.1.3 for a complete decision guide.

The canonical example is the **Lattice Deduction Transformer** (LDT, Davis et al., 2026): an 800K-parameter recurrent transformer that achieves **100% accuracy on Sudoku-Extreme** while frontier LLMs score **0%** — while remaining empirically sound: it returns a correct answer or abstains, never an incorrect one.(Davis et al., 2026)

***

## Sections in This Chapter

* [4.4.1.1 Abstract Interpretation Foundations](/neuro-symbolic-ai-in-practice/part-iii-core-approaches/chapter-4/4-4-pure-neural/4-4-differentiable-abstract-interpretation/4-4-foundations.md) — Galois connections, optimal deduction operator, connection to constraint propagation
* [4.4.1.2 The Lattice Deduction Transformer (LDT)](/neuro-symbolic-ai-in-practice/part-iii-core-approaches/chapter-4/4-4-pure-neural/4-4-differentiable-abstract-interpretation/4-4-foundations-2.md) — Architecture, lattice encoding, training procedure, and on-policy training
* [4.4.1.3 The Scale Inversion Result](/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) — Why 800K parameters beats frontier LLMs at 0%, and what this means for practitioners
* [4.4.1.4 Historical Lineage: SATNet to LDT](/neuro-symbolic-ai-in-practice/part-iii-core-approaches/chapter-4/4-4-pure-neural/4-4-differentiable-abstract-interpretation/4-4-design-principles-2.md) — Progression from SATNet through NeuroSAT, NeurASP, and Neural Abstract Interpretation to LDT
* [4.4.1.5 When to Use DAI](/neuro-symbolic-ai-in-practice/part-iii-core-approaches/chapter-4/4-4-pure-neural/4-4-differentiable-abstract-interpretation/4-4-design-principles-3.md) — Decision guide for the DAI paradigm

***

## References

1. Davis, Liam, Leopold Haller, Alberto Alfarano, and Mark Santolucito. "Lattice Deduction Transformers." *arXiv preprint* arXiv:2605.08605 (2026). <https://arxiv.org/abs/2605.08605>


---

# 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.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.