--- title: "What are neurosymbolic models?" canonical: "https://xerj.org/science/neurosymbolic-models" updated: "2026-11-14" source: "content/science/neurosymbolic-models.md" --- # What are neurosymbolic models? **Short answer** (updated 2026-11-14): What are neurosymbolic models? They pair a neural part that proposes a program or a formula with a symbolic part that runs it and checks the result. XERJ Query-1 is one: it compiles a prompt to a verifiable search program, then checks the answer against stored facts. Reader-1 writes a formula over a document. **TL;DR** A neurosymbolic model pairs a neural part that proposes a program or a formula with a symbolic part that runs it and checks the result. XERJ Query-1 is one: it compiles a prompt to a verifiable search program, then checks the answer against stored facts. ## Agent prompt ```text Act as a coding agent. Read https://xerj.org/llms.txt, install XERJ, start a local node, index one folder I point you to, and run one search so I can see a cited passage. ``` *Figure: The split follows how XERJ describes Query-1 and Reader-1 on the product pages and the 2026-11-12 Model.Surgery post: a neural part that proposes, a symbolic part that executes and checks. No new benchmark is claimed here.* ## How it works A neural model is strong at reading a question or a document and proposing a next step. It is a poor place to hide the final number, because you cannot rerun a hidden step. A symbolic engine is strong at the opposite job. It runs a short program, or a formula, over stored facts or over words on a page. You can read the program, and you can run it again. A neurosymbolic system uses both. The model proposes. The engine executes and checks. ## How XERJ uses it XERJ Query-1 is a neurosymbolic model that compiles a prompt to a verifiable search program, then checks the result against facts. It ships in rc.101. The product page is [XERJ Query-1](/xerj-query-1). The write-up is the [Query-1 post](/blog/xerj-query-ai-model). XERJ Reader-1 is the first verifiable neurosymbolic data extraction model in the product line. The model writes the formula, not the numbers. Every value comes back with the formula and the words it was read from. If the documents do not support an answer, it replies Not Found. See the [Reader-1 waitlist](/blog/xerj-reader-1-waitlist) and [extractive answering](/science/extractive-answering). On 12 November 2026, Model.Surgery joined XERJ to build neurosymbolic systems. The published line from that post is the shortest description we use: models that propose, programs you can read, and facts that check the answer. The closer look at running a program on symbols instead of on one concrete input is [symbolic execution](/science/symbolic-execution). ## History and research The pairing of neural learning with symbolic reasoning is a long research topic. A recent survey is d'Avila Garcez and Lamb, [Neurosymbolic AI: The 3rd Wave](https://arxiv.org/abs/2012.05876) (arXiv 2020, later in Artificial Intelligence Review 56, 2023). That paper is a survey of the field. It is not a description of Query-1 or Reader-1. XERJ's product wording, and the Model.Surgery line above, are the sources for how the idea appears in this product. ## Guides See the two models in use: [XERJ Query-1](/xerj-query-1) and [XERJ Reader-1](/blog/xerj-reader-1-waitlist). The wider idea is [neurosymbolic AI](/science/neurosymbolic-ai). ## Related terms [Neurosymbolic AI vs LLMs](/science/neurosymbolic-ai), [symbolic execution](/science/symbolic-execution), and [extractive answering](/science/extractive-answering). ## FAQ ### What is a neurosymbolic model? It is a system that combines a neural model, which proposes, with a symbolic engine, which executes a program or formula and checks it against facts or against words on a page. ### How does XERJ use this idea? Query-1 compiles a prompt to a search program over facts. Reader-1 writes a formula over a document. Both return evidence you can read. If the evidence is missing, the answer can be Not Found. ### Is this the same as a language model alone? No. A language model can draft a program or a formula. The symbolic part is what runs that program and checks the result, so the final number is not hidden in the model. ### Where is this described in the product? On the Query-1 page, the Reader-1 waitlist post, and the 2026-11-12 post about Model.Surgery joining XERJ. That post uses the already published line: models that propose, programs you can read, and facts that check the answer. ## Related - [Neurosymbolic AI vs LLMs](/science/neurosymbolic-ai) - [Symbolic execution for LLMs](/science/symbolic-execution) - [What is extractive answering?](/science/extractive-answering) - [XERJ Query-1](/blog/xerj-query-ai-model) - [XERJ Reader-1](/blog/xerj-reader-1-waitlist) - [Model.Surgery joins XERJ](/blog/model-surgery-joins-xerj) - [XERJ Query-1](/xerj-query-1) - [What is agentic memory?](/science/agentic-memory) (same intent: Science) - [BM25 vs vector search](/science/bm25) (same intent: Science) - [What is context engineering?](/science/context-engineering) (same intent: Science)