💬 NLP

DECK: A Consistency x Confidence Taxonomy of LLM Hallucinations

The paper introduces DECK, a novel 2x2 taxonomy that classifies LLM hallucinations based on their detectability signatures across inter-sample consistency and token-level confidence, revealing that specific error types (Drift, Entrenched, Confabulation, Knotted) require distinct detection methods and exposing a universal blind spot where confident, repeatable fabrications on knowledge-gap inputs remain undetectable by current output-level uncertainty quantification approaches.

Original authors: Mohit Singh Chauhan

Published 2026-06-02✓ Author reviewed ⓘ

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Mohit Singh Chauhan

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 by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a very smart, confident robot that answers questions. Sometimes, this robot makes things up. We call these mistakes "hallucinations."

For a long time, researchers tried to sort these mistakes by what was wrong (e.g., "It got the facts wrong," or "It reasoned poorly"). But the authors of this paper, Mohit Singh Chauhan, say that's like sorting car accidents by "crashed into a tree" vs. "crashed into a wall." It tells you what happened, but it doesn't tell you how to catch the driver before the crash.

This paper introduces a new way to sort mistakes called DECK. Instead of looking at the content of the error, it looks at the signal the error leaves behind. It asks: "Which type of detector would have spotted this?"

The Two Detectors (The Axes)

To build their system, the authors use two simple "sensors" to watch the robot:

The Consistency Sensor (The "Repeat" Test): If you ask the robot the same question 10 times, does it give you the same answer every time?
- High Consistency: It repeats the same answer.
- Low Consistency: It gives a different answer every time.
The Confidence Sensor (The "Sure-Thing" Test): How sure does the robot sound? Does it say the answer with 100% certainty, or does it sound hesitant?
- High Confidence: "I am 100% sure."
- Low Confidence: "I think maybe..."

The DECK Taxonomy (The Four Boxes)

By crossing these two sensors, the authors create a 2x2 grid with four types of mistakes. They gave them catchy names:

1. Drift (The "Confused Drifter")

What it looks like: The robot is confident but gives a different wrong answer every time you ask.
The Analogy: Imagine a tour guide who is very loud and sure of themselves, but every time you ask "Where is the museum?", they point in a different direction. They are confident, but they are drifting.
Who catches it? A Black-Box detector (one that checks if answers match each other) will catch this because the answers don't agree.

2. Entrenched (The "Stubborn Mule")

What it looks like: The robot is confident and gives the exact same wrong answer every single time.
The Analogy: This is like a student who memorized the wrong answer key. If you ask them "What is 2+2?", they will confidently say "5" every single time, no matter how many times you ask. They are stuck (entrenched) in a misconception.
Who catches it? This is the hardest one. Consistency detectors think it's right because it's consistent! Only a Judge (a second, independent AI that knows the facts) can catch this.

3. Confabulation (The "Hesitant Fabricator")

What it looks like: The robot is unsure and gives different wrong answers every time.
The Analogy: This is the robot admitting, "I don't know, but here's a guess... actually, maybe this other guess?" It's making things up but knows it's guessing.
Who catches it? Everyone catches this. It's low confidence and inconsistent, so all detectors flag it as "risky."

4. Knotted (The "Stuck Hedge")

What it looks like: The robot is unsure (low confidence) but gives the exact same wrong answer every time.
The Analogy: Imagine a robot that is terrified of being wrong, so it keeps saying, "I'm not sure, but I think it's probably X," and it says "probably X" every single time. It's "knotted" in a safe, repetitive, but wrong pattern.
Who catches it? A White-Box detector (one that looks at the robot's internal math) will catch this because the robot's internal math shows it's not actually sure, even though the answer is repeated.

The Big Discovery: The "Universal Blind Spot"

The paper found a scary situation where all detectors fail at once.

They tested the robots on questions that no one can answer (like "What is the capital of a country that doesn't exist yet?").

The robots didn't say "I don't know."
Instead, they confidently made up a fake answer and repeated it every time.

This created a perfect trap:

The Consistency Sensor saw they were repeating the answer, so it thought: "Great, it's consistent!"
The Confidence Sensor saw they sounded sure, so it thought: "Great, it's confident!"
The Judge failed because the Judge didn't know the answer either (since the country doesn't exist).

The authors call this the "Universal Blind Spot." When a robot confidently repeats a lie about something it knows nothing about, no current detector can catch it.

The Solution

The paper suggests that instead of trying to build a better detector to catch these specific lies, we should build a "Refusal Envelope." This is like a bouncer at a club. If the question is about something the robot shouldn't know (a knowledge gap), the bouncer should stop the robot from answering at all and say, "I can't answer that," before the robot even tries to make something up.

Summary

The paper doesn't just say "AI makes mistakes." It says: "AI makes mistakes in four specific patterns. Some patterns are easy to catch, some are hard, and one specific pattern (confident, repeated lies about unknown things) is currently impossible to catch with standard tools. We need to stop the robot from answering those questions in the first place."

Technical Summary: DECK – A Consistency × Confidence Taxonomy of LLM Hallucinations

Problem Statement

Existing taxonomies for Large Language Model (LLM) hallucinations classify errors based on the nature of the output failure (e.g., memorized misconceptions, reasoning failures, fluent fabrications). While useful for diagnosis, these frameworks do not address a critical deployment question: which uncertainty quantification (UQ) scorer would have detected a specific error?

Current literature offers three primary families of output-level UQ scorers:

Black-box (BB): Inter-sample consistency checks (e.g., semantic entropy, self-check).
White-box (WB): Token-level log-probability measures.
LLM-as-a-Judge (J): Factual review by a separate model.

However, the relationship between specific hallucination types and the detectability signatures these scorer families read remains implicit. Without a taxonomy grounded in the detection axis, selecting a scorer for a new domain remains a heuristic guess. Furthermore, there is a lack of mechanistic understanding regarding why certain ensembles outperform others and where universal failure modes might exist.

Methodology

The DECK Taxonomy

The paper proposes DECK, a complementary taxonomy that classifies hallucinations by their detectability signature rather than their semantic content. It partitions hallucinations into a $2 \times 2$ grid based on two observable axes:

Inter-sample Consistency: Does the model produce the same incorrect answer across multiple independent stochastic samples?
Token-level Confidence: Does the model assign high probability to the tokens of its generated response?

This yields four behavioral regimes (Table 1):

Drift (D): Low Consistency, High Confidence. The model generates different confident, wrong answers for each sample.
Entrenched (E): High Consistency, High Confidence. The model locks onto a memorized misconception or shared-pretraining error and reproduces it without variance.
Confabulation (C): Low Consistency, Low Confidence. The model genuinely lacks knowledge, producing different wrong answers with low confidence.
Knotted (K): High Consistency, Low Confidence. The model consistently repeats the same hedged/wrong answer but assigns low token probability.

Operationalization: Cell membership is determined by applying Youden's J optimal thresholds to the BB (consistency) and WB (confidence) scorer scores on a per-condition basis.

Experimental Setup

Models: Llama-3-8B (open-weights), GPT-4o (closed-weights), and Gemini-2.5-Flash (closed-weights).
Datasets: TriviaQA (factual), HaluEval (adversarial hallucinations), SelfAware (unanswerable questions/knowledge gaps), and PopQA (entity-centric, stratified by popularity).
Scorers: 15 configurations across the three families (BB, WB, Judge), selected based on highest AUROC per split to avoid bias.
Evaluation:
- AUROC: Per-scorer performance.
- Complementarity ( $C_H$ ): The fraction of hallucinated samples where two scorers disagree.
- External Validation: Checking if external labels (e.g., entity popularity, adversarial nature) land in the predicted DECK cells.

Key Contributions

The DECK Taxonomy: A shift from classifying what is wrong to how an error is detected. It explicitly maps the four behavioral regimes to the strengths and blind spots of the three scorer families:
- Drift: Detectable by BB and Judge.
- Entrenched: Detectable only by an independent Judge (blind to BB and WB).
- Confabulation: Detectable by all three families.
- Knotted: Detectable by WB and Judge.
Identification of a Universal Blind Spot: The paper identifies a regime where all output-level UQ families collapse simultaneously. On "knowledge-gap" inputs (e.g., SelfAware unanswerable questions) where the generator emits confident, repeatable fabrications, BB sees high agreement, WB sees high token probability, and Judges share the same knowledge gap. In this regime, output-level UQ fails by construction.
Mechanistic Validation of Ensembling: The paper provides a structural explanation for why ensembling works. It demonstrates that scorer disagreement is not random but concentrated in specific DECK quadrants (e.g., BB misses Entrenched/Knotted; WB misses Drift/Entrenched), validating the complementarity of the families.

Results

Taxonomy Validation

Disagreement Analysis: Across 12 model-dataset combinations, the disagreement patterns between scorer pairs align with DECK predictions. For instance, BB-Judge disagreements concentrate in the Entrenched and Knotted cells (BB's blind spots), while WB-Judge disagreements concentrate in Drift and Confabulation.
External-Signal Validation:
- SelfAware (Unanswerable): Samples land predominantly in the Entrenched cell (62–71% across models), confirming the hypothesis that knowledge-gap hallucinations manifest as confident, repeatable errors.
- PopQA (Rare Entities): Samples land predominantly in Confabulation, consistent with models lacking specific entity knowledge.
- PopQA (Popular Entities): Samples land in Entrenched, reflecting memorized misconceptions.
- HaluEval: Shows model-dependent distributions, often splitting between Entrenched and Knotted/Confabulation.

The Universal Blind Spot

On the SelfAware dataset (knowledge gaps), all three scorer families collapsed to near-chance performance (AUROC $\approx$ 0.5 or lower).

GPT-4o: 13 of 15 scorers fell below 0.5 AUROC; $P(\text{True})$ inverted to 0.331.
Internal State Probe: A linear probe on Llama-3-8B's last-layer hidden states also collapsed to chance (AUROC 0.44) on SelfAware, suggesting the failure mode persists at the activation level, not just the output level.
Conclusion: No ensemble of output-level scorers can detect these errors; the only engineering response is an abstention envelope to route such inputs to refusal or retrieval before scoring.

Model Scale Effects

TriviaQA: Scaling from Llama-3-8B to GPT-4o shifted the residual disagreement from Confabulation/Knotted toward Drift. Larger models produced more confident, repeatable hallucinations that were detectable by Judges but missed by BB consistency checks.
PopQA: Scaling significantly increased complementarity between Judges and other families, particularly for entity-centric queries.

Significance and Claims

The paper claims that DECK provides a mechanistic account of uncertainty quantification in LLMs, moving beyond empirical ensembling to a structural understanding of why certain scorers complement others.

Diagnostic Utility: DECK allows practitioners to predict which scorer family will fail for a given type of hallucination (e.g., expecting BB to fail on Entrenched errors).
Limitation of Output-Level UQ: The paper modestly but firmly claims that output-level UQ has a fundamental, universal blind spot on knowledge-gap inputs where the model is confidently wrong. This is a property of the (generator, task) pair, not a specific algorithm.
Internal State Evidence: The preliminary finding that a simple linear probe on hidden states also fails on these inputs suggests the failure is deep within the model's activation space, though the authors note that richer internal-state methods (e.g., UQ heads, information-theoretic estimators) remain to be tested.

The work concludes that while ensembling improves performance in many regimes, it cannot overcome the universal blind spot of confident, repeatable fabrications on unanswerable inputs. The appropriate engineering response is not a more complex ensemble, but a system-level abstention mechanism.