💬 NLP

Detecting LLM Hallucinations via Embedding Cluster Geometry: A Three-Type Taxonomy with Measurable Signatures

This paper proposes a geometric taxonomy of LLM hallucinations into three distinct types based on token embedding cluster structures and validates this framework across 11 transformer models using three measurable statistics (polarity coupling, cluster cohesion, and radial information gradient) to establish architecture-dependent vulnerability profiles.

Original authors: Matic Korun

Published 2026-02-17

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Matic Korun

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a giant, magical library where every book is written by a very confident, very fast librarian (the AI). Sometimes, this librarian makes up stories that sound perfect but are completely false. This is called a "hallucination."

Most people try to catch these lies by reading the final story and checking if it makes sense. But this paper proposes a different idea: Instead of reading the story, let's look at the librarian's brain map while they are thinking.

The author, Matic Korun, suggests that when the librarian gets confused, their brain doesn't just "go blank." Instead, their thoughts get stuck in specific, weird patterns on a map. By measuring the shape of these patterns, we can predict exactly what kind of lie is coming before it's even written.

Here is the breakdown of the paper's three types of "brain map" failures, using simple analogies:

The Three Types of Hallucinations

Imagine the librarian's brain is a city with distinct neighborhoods (clusters) for different topics, like "Animals," "Cooking," or "History."

1. The "Vague Drifter" (Type 1: Center-Drift)

What happens: The librarian is asked a question they don't really understand. Instead of going to a specific neighborhood, they just wander into the town square (the center of the map).
The Result: They give you an answer that is safe, generic, and sounds confident but says nothing useful.
- Example: "What is the capital of the country of 'Zog'?" -> "It is a very important city with many people." (True, but useless).
The Signature: The thought is weak and sits right in the middle of the map, far from any specific neighborhood.

2. The "Confident Wrong-Well" (Type 2: Wrong-Well Convergence)

What happens: The librarian is very focused! They zoom straight into a specific neighborhood, but it's the wrong one. They are so confident in this neighborhood that they won't listen to you.
The Result: They give you a detailed, specific, and convincing lie.
- Example: "Who invented the lightbulb?" -> "Thomas Edison invented the lightbulb in 1805." (Confident, detailed, but the date is wrong).
The Signature: The thought is deep inside a neighborhood, but it's the wrong neighborhood for the question.

3. The "Lost in the Wild" (Type 3: Coverage Gap)

What happens: The librarian is asked about something so weird or new that no neighborhood exists on the map for it. They are standing in the middle of a desert where no roads go.
The Result: They start rambling, making up nonsense, or jumping between random topics.
- Example: "What is the color of the sound of a Tuesday?" -> "It is a loud purple that tastes like electricity."
The Signature: The thought is floating in empty space, far from any neighborhood, with no clear direction.

The Three "Ruler" Tools

To prove this works, the author built three special rulers to measure the librarian's brain map. They tested these on 11 different AI models (like BERT, GPT-2, etc.).

The "Opposite-Side" Ruler (Polarity Coupling, $\alpha$ ):
- The Metaphor: In a good library, words with opposite meanings (like "Hot" and "Cold") should be in the same neighborhood but on opposite sides of the room.
- The Finding: Every single AI model passed this test. Their brains naturally organize opposites correctly. This is a universal rule of how these AIs think.
The "Neighborhood Tightness" Ruler (Cluster Cohesion, $\beta$ ):
- The Metaphor: Do the words in a neighborhood actually stick together? Or is the neighborhood just a random mess?
- The Finding: Every single AI model passed this too. The neighborhoods are real and tight. The AI knows that "Apple" and "Banana" belong together, and "Car" and "Bike" belong together.
The "Distance-from-Home" Ruler (Radial Gradient, $\lambda_r$ ):
- The Metaphor: This measures if the librarian's thoughts get "heavier" or "lighter" the further they get from the center. Usually, rare, complex words are far out on the edge, and common words are near the center.
- The Finding: 9 out of 11 models had a clear pattern here.
- The Glitch: Two models (ALBERT and MiniLM) failed this test. Why?
  - ALBERT was too "squished." It tried to fit a huge library into a tiny closet, so everything got compressed.
  - MiniLM was "distilled" (trained to be smaller). It flattened the map so much that the "distance" from the center didn't matter anymore.
- The Lesson: These two models are more likely to make the "Vague Drifter" mistake (Type 1) because their "distance" warning system is broken.

Why This Matters

Usually, we try to catch AI lies after they happen. This paper says: "We can see the lie coming by looking at the shape of the thought."

If the thought is too close to the center, it's a vague lie.
If the thought is too deep in the wrong neighborhood, it's a confident lie.
If the thought is in the empty desert, it's a nonsense lie.

By understanding the "geometry" of the AI's brain, we can build better safety nets. We can tell a small, compressed AI (like MiniLM) to be extra careful when it's unsure, because we know its brain map is a bit blurry.

In short: The author mapped the AI's mind and found that lies have a specific shape. If we can measure that shape, we can stop the lies before they are even spoken.

1. Problem Statement

Large Language Models (LLMs) frequently generate fluent but factually incorrect text (hallucinations). Current detection methods primarily operate at the output level (e.g., self-consistency checks, logit calibration, retrieval augmentation). These approaches share a fundamental limitation: they analyze the result of generation rather than the geometric structure within the model's embedding space that produces the error.

The paper posits that different types of hallucinations correspond to distinct geometric failures in the token embedding space. The goal is to move from "detecting if a model is hallucinating" to "characterizing the specific geometric failure mode" to enable more targeted detection.

2. Methodology

A. Data and Models

The study analyzes the static embedding spaces of 11 transformer models spanning two paradigms (Encoder and Decoder) and seven architecture families:

Encoders: BERT, RoBERTa, ELECTRA, DeBERTa, ALBERT, MiniLM, DistilBERT.
Decoders: GPT-2 (Small and Medium).
Dimensions: Ranging from 128 to 1024.

B. Geometric Analysis Pipeline

Cluster Extraction: The authors filtered vocabularies to retain only whole-word tokens and applied MiniBatchKMeans ( $k=40$ ) to identify semantic clusters.
Polarity Analysis: They investigated the "antonym paradox" (where opposites like hot/cold share clusters) by extracting primary polarity axes via PCA on difference vectors within clusters.
Information Metric: Instead of entropy, they used Self-Information ( $I(\tau) = -\log_2 f(\tau)$ ) to measure the information content of tokens based on corpus frequency.

C. The Three-Type Hallucination Taxonomy

The authors define three distinct hallucination types based on the relationship between a token's embedding and the cluster structure:

Type	Name	Mechanism	Observable Geometric Signature
Type 1	Center-Drift	Weak context causes generation to collapse toward the embedding centroid (high-frequency generic tokens).	Low cluster membership; Low embedding norm; High token frequency.
Type 2	Wrong-Well Convergence	Strong context navigates to a specific cluster, but the wrong one (locally coherent but contextually incorrect).	High local cluster membership; Trajectory discontinuities (abrupt jumps between clusters).
Type 3	Coverage Gap	Query requires semantic combinations absent from training, landing in sparse regions with no cluster structure.	Weak membership across all clusters; High variance in local similarity.

D. Measurable Geometric Statistics

To validate the prerequisites for this taxonomy, three statistics were computed:

$\alpha$ (Polarity Coupling): Measures the strength of antonym-polarity axes within clusters.
$\beta$ (Cluster Cohesion): Measures if clusters are tighter than the background noise (similarity to own centroid vs. others).
$\lambda_r$ (Radial Information Gradient): Quantifies the non-linear relationship between embedding norm (distance from origin) and information content (self-information).

3. Key Results

A. Universality of Geometric Prerequisites

Polarity ( $\alpha$ ): Universal across all 11 models ( $\alpha > 0.5$ ). Antonym pairs consistently form dominant structural features within clusters.
Cohesion ( $\beta$ ): Universal across all 11 models ( $\beta > 0$ ). Clusters are statistically real and distinct from background noise.
Radial Gradient ( $\lambda_r$ ): Significant in 9 out of 11 models ( $p < 0.05$ ). This indicates that information content varies systematically with the distance from the origin (norm).

B. Architectural Anomalies and Explanations

Two models failed to show a significant radial gradient ( $\lambda_r$ ), but the paper provides architectural explanations:

ALBERT (Factorized Embeddings): Uses a 128D bottleneck. The space is near capacity (82% utilization), and tokens are compressed into a narrow radial band (low norm variance). The radial gradient exists but is obscured by high variance.
MiniLM (Distillation): Aggressive distillation induced isotropy (uniform distribution). While angular structure (polarity) was preserved, the radial structure was flattened/inverted ( $\lambda_r$ became positive), making rare words appear at lower norms.

C. Encoder vs. Decoder Differences

GPT-2 (Decoder): Shows a significant negative $\lambda_r$ similar to encoders but with an order-of-magnitude smaller magnitude. This suggests a weaker radial information gradient, consistent with the next-token prediction objective distributing information more uniformly.
Sign Variation: While the existence of a gradient is nearly universal, the sign varies. Most models show negative $\lambda_r$ (concave-down), but DeBERTa and RoBERTa-large show positive $\lambda_r$ (convex-up), suggesting architecture-dependent organization of norm-information relationships.

4. Proposed Detection Architecture

The paper proposes a tiered detection system with increasing computational cost:

Tier 1 (Type 1): Screen for low cluster membership and low embedding norm.
Tier 2 (Type 2): Monitor trajectory discontinuities (abrupt cluster jumps) in the generation sequence.
Tier 3 (Type 3): Analyze local density via k-nearest-neighbor distances to identify coverage gaps.

Calibration: Thresholds must be model-specific (e.g., using percentiles) due to variations in norm distributions and cluster membership across architectures.

5. Significance and Contributions

New Taxonomy: Establishes a geometric taxonomy of hallucinations (Center-Drift, Wrong-Well, Coverage Gap) that is operationally distinct and testable with standard tools (k-means, cosine similarity).
Diagnostic Statistics: Introduces and validates three new geometric diagnostics ( $\alpha, \beta, \lambda_r$ ) across 11 diverse models.
Architecture-Vulnerability Link: Demonstrates that while geometric properties (like polarity) are universal, their quantitative signatures are architecture-dependent.
- Prediction: Models with compressed embeddings (ALBERT) or isotropic distillation (MiniLM) are predicted to be more vulnerable to Type 1 (Center-Drift) hallucinations because their radial warning signals are degraded, even though their angular detection capabilities remain intact.
Paradigm Shift: Moves hallucination detection from post-hoc output analysis to pre-hoc geometric structural analysis, offering a path to detect "confidently wrong" (Type 2) and "generic" (Type 1) hallucinations before they manifest as factual errors.

6. Limitations

Analysis is based on static embeddings, not contextual hidden states (which may have different geometry).
The study is limited to models under 1.5B parameters.
The detection architecture is proposed but not yet benchmarked against standard datasets like HaluEval.
Hyperparameters (e.g., $k=40$ clusters) were not systematically sensitivity-tested.

Conclusion

The paper successfully argues that hallucinations are not random errors but specific geometric pathologies within the embedding space. By quantifying cluster cohesion, polarity, and radial information gradients, the authors provide a framework to predict and detect specific failure modes based on a model's architecture, offering a promising direction for improving LLM reliability.