The Hallucination That Looks Identical — Until You Read It Twice

2026-01-11 · Source: Agus’s Substack · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, long

Summary

This article, part three of a series on geometric approaches to typed hallucination detection, introduces a method for identifying "Type III" relational inversion hallucinations in Large Language Models (LLMs). These errors involve swapping subject-object roles, such as "The hunter shot the bear" becoming "The bear shot the hunter," which are difficult for standard fluency or similarity checks to catch. The proposed solution moves beyond a naive triplet extraction and bivector approach, which fails on passive voice, to utilize cross-encoder Natural Language Inference (NLI) models. By analyzing the hidden state geometry of these models, the method distinguishes between inversions and negations using subspace projection, achieving 100% accuracy on benchmarks. This technique provides a continuous, calibratable score for inversion confidence, crucial for audit trails and regulatory compliance in applications like RAG pipelines.

Key takeaway

For AI Engineers developing RAG systems or other LLM applications, understanding and implementing Type III hallucination detection is critical for robust, compliant systems. Your team should adopt subspace projection methods using cross-encoder NLI models to accurately distinguish relational inversions from negations, especially in regulated domains like finance or medicine. This approach provides a precise, auditable classification of "who did what to whom" errors, enhancing model reliability and reducing liability.

Key insights

Subspace projection on cross-encoder hidden states effectively detects relational inversions, even with passive voice.

Principles

• Relational inversions are a distinct hallucination type.
• Cross-encoders internalize passive voice semantics.
• Subspaces capture subtle geometric differences better than single directions.

Method

The method uses a cross-encoder NLI model to identify contradictions, then projects the pooled hidden state onto pre-calibrated inversion and negation subspaces to classify the specific error type, providing a continuous confidence score.

In practice

• Integrate Type III detection into RAG audit loops.
• Use cross-encoder pooled embeddings for geometric analysis.
• Build orthonormal bases from calibration examples for subspaces.

Topics

• LLM Hallucination Detection
• Relational Inversion
• Cross-Encoder NLI
• Subspace Projection
• Geometric AI

Best for: AI Engineer, Machine Learning Engineer, AI Researcher

Related on AIssential

• Type VI Hallucination: When LLMs Apply the Wrong Domain’s Rules — Type VI hallucination detection identifies domain-specific semantic mismatches in LLM outputs using geometric embedding analysis.
• Learning Uncertainty from Sequential Internal Dispersion in Large Language Models — SIVR detects LLM hallucinations by analyzing internal representation variance across layers and tokens.
• Hallucinations in LLMs: A Deep Technical Dive into Causes, Detection, and Mitigation — LLM hallucinations are inherent to their probabilistic nature, driven by training objectives that prioritize fluency over factual accuracy.
• Hallucinations as Orthogonal Noise: Inference-Time Manifold Alignment via Dynamic Contextual Orthogonalization — Hallucinations in LLMs are modeled as orthogonal noise, which Dynamic Contextual Orthogonalization (DCO) suppresses by aligning latent representations.
• Steering the Verifiability of Multimodal AI Hallucinations — Multimodal AI hallucinations vary in human verifiability, requiring distinct intervention strategies for obvious versus elusive errors.
• Detecting Hallucinations in SpeechLLMs at Inference Time Using Attention Maps — Attention-derived metrics enable efficient, inference-time hallucination detection in SpeechLLMs without gold-standard outputs.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Agus’s Substack.