Marcelo Labre developed a neuro-symbolic pipeline using OpenMath ontology to enhance language model reliability
Research shows ontology-guided context improves performance when retrieval quality is high
Irrelevant context actively degrades model performance, highlighting challenges of neuro-symbolic approaches
The study demonstrates how structured domain knowledge can reduce hallucinations and brittleness in AI systems
📖 Full Retelling
Researcher Marcelo Labre published a paper on arXiv on February 19, 2026, investigating how formal domain ontologies can enhance language model reliability through retrieval-augmented generation, addressing fundamental limitations in AI systems that affect high-stakes specialist fields requiring verifiable reasoning. The paper introduces a neuro-symbolic pipeline that leverages the OpenMath ontology with hybrid retrieval and cross-encoder reranking to inject relevant mathematical definitions into model prompts, aiming to reduce hallucinations and brittleness in language models. Using mathematics as a proof of concept, Labre evaluated this approach on the MATH benchmark with three open-source models, revealing that while ontology-guided context improves performance when retrieval quality is high, irrelevant context actively degrades it. This finding highlights both the promise and challenges of neuro-symbolic approaches to AI development, suggesting that while domain knowledge can enhance language model performance in specialized fields, the quality and relevance of that knowledge are critical factors for success.
🏷️ Themes
AI reliability, Neuro-symbolic integration, Domain knowledge grounding
In the field of artificial intelligence (AI), a hallucination or artificial hallucination (also called bullshitting, confabulation, or delusion) is a response generated by AI that contains false or misleading information presented as fact. This term draws a loose analogy with human psychology, where...
A language model is a computational model that predicts sequences in natural language. Language models are useful for a variety of tasks, including speech recognition, machine translation, natural language generation (generating more human-like text), optical character recognition, route optimizati...
This research addresses critical limitations of language models like hallucination and lack of formal grounding, which is especially important for high-stakes fields requiring verifiable reasoning. By demonstrating a neuro-symbolic approach using mathematical ontologies, it offers a pathway to improve AI reliability and precision in specialized domains.
Context & Background
Language models often struggle with factual accuracy and formal reasoning
Neuro-symbolic AI combines neural networks with symbolic systems for better reasoning
Mathematical ontologies like OpenMath provide structured domain knowledge
Retrieval-augmented generation uses external knowledge to enhance model outputs
What Happens Next
The paper will likely undergo peer review for the NeuS 2026 conference. Further research may explore applying this ontology-guided approach to other technical domains beyond mathematics to validate its broader utility.
Frequently Asked Questions
What problem does this research solve?
It tackles language model limitations like hallucination and brittleness in mathematical reasoning by using formal ontologies.
What was the key finding of the study?
Ontology-guided context improves performance when retrieval is accurate, but irrelevant context can degrade results.
Which benchmark was used for evaluation?
The research was evaluated using the MATH benchmark with open-source language models.
Original Source
--> Computer Science > Artificial Intelligence arXiv:2602.17826 [Submitted on 19 Feb 2026] Title: Ontology-Guided Neuro-Symbolic Inference: Grounding Language Models with Mathematical Domain Knowledge Authors: Marcelo Labre View a PDF of the paper titled Ontology-Guided Neuro-Symbolic Inference: Grounding Language Models with Mathematical Domain Knowledge, by Marcelo Labre View PDF HTML Abstract: Language models exhibit fundamental limitations -- hallucination, brittleness, and lack of formal grounding -- that are particularly problematic in high-stakes specialist fields requiring verifiable reasoning. I investigate whether formal domain ontologies can enhance language model reliability through retrieval-augmented generation. Using mathematics as proof of concept, I implement a neuro-symbolic pipeline leveraging the OpenMath ontology with hybrid retrieval and cross-encoder reranking to inject relevant definitions into model prompts. Evaluation on the MATH benchmark with three open-source models reveals that ontology-guided context improves performance when retrieval quality is high, but irrelevant context actively degrades it -- highlighting both the promise and challenges of neuro-symbolic approaches. Comments: Submitted to NeuS 2026. Supplementary materials and code: this https URL Subjects: Artificial Intelligence (cs.AI) ; Machine Learning (cs.LG); Symbolic Computation (cs.SC) Cite as: arXiv:2602.17826 [cs.AI] (or arXiv:2602.17826v1 [cs.AI] for this version) https://doi.org/10.48550/arXiv.2602.17826 Focus to learn more arXiv-issued DOI via DataCite (pending registration) Submission history From: Marcelo Labre [ view email ] [v1] Thu, 19 Feb 2026 20:45:16 UTC (2,970 KB) Full-text links: Access Paper: View a PDF of the paper titled Ontology-Guided Neuro-Symbolic Inference: Grounding Language Models with Mathematical Domain Knowledge, by Marcelo Labre View PDF HTML TeX Source view license Current browse context: cs.AI < prev | next > new | recent | 2026-02 Change t...
