Real-Time Hallucination Detection in Large Language Models via Topological Strain and Geometric Homeostasis
Fuente:
Zenodo
Enregistré dans:
| Auteur principal: | |
|---|---|
| Format: | Recurso digital |
| Publié: |
Zenodo
2026
|
| Sujets: | |
| Accès en ligne: | |
| Tags: |
Ajouter un tag
Pas de tags, Soyez le premier à ajouter un tag!
|
| _version_ | 1866902308331192320 |
|---|---|
| author | Teixido, John |
| author_facet | Teixido, John |
| contents | <p>The tendency of Large Language Models (LLMs) to confidently generate factually incorrect information—termed “hallucination”—remains a critical barrier to enterprise deployment. Current mitigation strategies rely on semantic evaluation or external retrieval (RAG), which are computationally expensive and post-hoc. We propose a fundamental, pre-generation solution grounded in Topological Analytical Homeostasis (TAH). We introduce a method for calculating the Topological Strain of a generated token vector relative to the machine-verified Teixido Envelope (τ = −0.5). By measuring the geometric deviation of the activation manifold in real-time, the system provides a deterministic, mathematical confidence score before the token is rendered to the user. We detail the implementation of this ”Hallucination Meter” and the Teixido Inhibitory Gating (TIG) mechanism, which physically intercepts and suppresses out-of-distribution confabulations at the logic level.</p> |
| format | Recurso digital |
| id | zenodo_https___doi_org_10_5281_zenodo_19652657 |
| institution | Zenodo |
| language | |
| publishDate | 2026 |
| publisher | Zenodo |
| record_format | zenodo |
| spellingShingle | Real-Time Hallucination Detection in Large Language Models via Topological Strain and Geometric Homeostasis Teixido, John LLM Hallucination Topological Strain AI Safety Teixido Envelope Contextual Monitoring Zero-MAC <p>The tendency of Large Language Models (LLMs) to confidently generate factually incorrect information—termed “hallucination”—remains a critical barrier to enterprise deployment. Current mitigation strategies rely on semantic evaluation or external retrieval (RAG), which are computationally expensive and post-hoc. We propose a fundamental, pre-generation solution grounded in Topological Analytical Homeostasis (TAH). We introduce a method for calculating the Topological Strain of a generated token vector relative to the machine-verified Teixido Envelope (τ = −0.5). By measuring the geometric deviation of the activation manifold in real-time, the system provides a deterministic, mathematical confidence score before the token is rendered to the user. We detail the implementation of this ”Hallucination Meter” and the Teixido Inhibitory Gating (TIG) mechanism, which physically intercepts and suppresses out-of-distribution confabulations at the logic level.</p> |
| title | Real-Time Hallucination Detection in Large Language Models via Topological Strain and Geometric Homeostasis |
| topic | LLM Hallucination Topological Strain AI Safety Teixido Envelope Contextual Monitoring Zero-MAC |
| url | https://doi.org/10.5281/zenodo.19652657 |