🤖 AI

Entropy Distribution as a Fingerprint for Hallucinations in Generative Models

This paper introduces the Calibrated Entropy Score (CES), a lightweight, single-pass black-box method that leverages token-level entropy distributions to detect LLM hallucinations with formal error guarantees and performance comparable to computationally expensive multi-sample approaches.

Original authors: Mattia J. Villani, Pranav Deshpande, Akshay Seshadri, Romina Yalovetzky, Niraj Kumar

Published 2026-05-28

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Mattia J. Villani, Pranav Deshpande, Akshay Seshadri, Romina Yalovetzky, Niraj Kumar

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 are listening to a storyteller. Sometimes, they tell a story that is perfectly true and flows naturally. Other times, they start "hallucinating"—making up facts, getting confused, or rambling in a way that doesn't make sense.

For a long time, computers trying to catch these lies had to do one of two things: either ask the storyteller to repeat the story five or six times to see if the details match (which is slow and expensive), or peek inside the storyteller's brain to see their secret notes (which is often impossible with closed systems).

This paper introduces a new, clever way to catch a liar that only requires listening to the story once. It's called the Calibrated Entropy Score (CES).

Here is how it works, using simple analogies:

1. The "Confidence Meter" (Entropy)

Every time a Large Language Model (LLM) writes a word, it has a "confidence meter" for what comes next.

High Confidence: The model knows exactly what to say. It's like a chef who knows the recipe perfectly. The "entropy" (a measure of uncertainty) is low.
Low Confidence: The model is guessing. It's like the chef staring at an empty fridge, unsure what to cook. The "entropy" is high.

2. The Old Way vs. The New Way

The Old Way (Perplexity):
Previous methods looked at the average confidence of the whole story.

Analogy: Imagine a student taking a test. The old method just looks at their average grade. If the average is 70%, they might pass. But this hides the truth: Did they get every question right with a 70% score? Or did they get 10 questions perfect (100%) and 10 questions completely wrong (0%)? The average is the same, but the second student is much more unstable.

The New Way (CES):
The authors realized that the shape of the confidence meter matters more than the average.

Analogy: When a model is telling the truth, its confidence meter is steady and predictable. When it starts hallucinating, the meter goes wild. It might be calm for a while, then suddenly spike into total confusion, then spike again.
The paper claims that this "wild pattern" is a fingerprint of a lie. Even if the average confidence looks normal, the spikes (the maximum uncertainty) and the overall shape of the curve give it away.

3. How CES Works (The "Fingerprint" Check)

The method works in two steps:

Learning the "Truth" Pattern: First, the system listens to many stories that are known to be true. It builds a "reference map" (a statistical curve) of what a truthful story's confidence meter looks like. It notes: "Okay, truthful stories usually have small bumps here, but rarely huge spikes there."
The One-Time Test: When a new story comes in, the system listens to it just once. It doesn't ask for a second opinion. It checks the story's confidence meter against the "Truth Map."
- If the story has a weird spike or a shape that doesn't fit the map, the system flags it: "This looks like a hallucination!"

4. Why This is a Big Deal

Speed: It only needs one pass through the text. It's like catching a liar by listening to them speak once, rather than making them repeat the story five times.
Black Box Friendly: You don't need to see the model's internal code or hidden thoughts. You just need the "logits" (the raw confidence numbers the model outputs before it picks a word), which most APIs provide.
Statistical Proof: The authors didn't just guess this works. They used math to prove that as the story gets longer, the chance of missing a lie drops to almost zero, and the chance of falsely accusing a truthful story also drops to almost zero.

Summary

Think of CES as a lie detector that doesn't need a polygraph machine or a second interview. It just listens to the rhythm of uncertainty in the text. If the rhythm is steady, it's likely true. If the rhythm has erratic, wild jumps (even if the average looks okay), it's likely a hallucination.

The paper tested this on 10 different AI models and 8 different types of questions (from math to trivia) and found that this "one-pass" method catches lies just as well as the expensive, multi-pass methods, but much faster and cheaper.

Technical Summary: Entropy Distribution as a Fingerprint for Hallucinations in Generative Models

Problem Statement

Large Language Models (LLMs) frequently generate factually incorrect outputs, a phenomenon known as hallucination, which undermines trust and limits deployment in high-stakes domains. Existing detection methods suffer from significant limitations:

Consistency-based methods (e.g., Semantic Entropy, SelfCheckGPT) require multiple forward passes ( $K \ge 5$ ), incurring high computational costs.
Internal-state methods require access to hidden representations or attention matrices, which is impossible with black-box API models.
Single-pass heuristics (e.g., Perplexity, Length-Normalized Entropy) reduce the full entropy signal to a single scalar summary, discarding rich distributional information that may be critical for detection.

The authors posit that the distribution of token-level entropies, rather than just their mean, serves as a unique "fingerprint" of hallucination.

Methodology: Calibrated Entropy Score (CES)

The paper proposes a framework that formalizes hallucination detection as a statistical hypothesis test, requiring only a single forward pass and black-box access to token logits.

1. Theoretical Foundation

The core hypothesis is that when a model hallucinates, the distribution of token entropies diverges from the distribution observed during faithful generations. This divergence is not merely a shift in the mean; the shape and tail behavior of the distribution carry independent signal.

Null Hypothesis ( $H_0$ ): The generated entropy sequence is consistent with a reference Cumulative Distribution Function (CDF), $F_0$ , derived from non-hallucinated data.
Alternative Hypothesis ( $H_1$ ): The sequence is inconsistent with $F_0$ , indicating hallucination.

2. The Calibrated Entropy Score (CES)

CES is a lightweight algorithm that combines two summary statistics of the entropy sequence:

Mean Entropy ( $\bar{h}$ ): Captures the central tendency.
Maximum Entropy ( $h_{max}$ ): Captures tail anomalies (high uncertainty spikes).

The score is computed as the geometric mean of the reference CDF values at these points:
$\text{CES}(h) = \sqrt{\hat{F}_0(\bar{h}) \cdot \hat{F}_0(h_{max})}$
where $\hat{F}_0$ is the empirical CDF estimated offline from a calibration set. A generation is flagged as a hallucination if the CES exceeds a calibrated threshold.

3. Calibration and Robustness

Supervised Calibration: Uses an oracle (e.g., human expert or LLM judge) to filter calibration data, constructing $\hat{F}_0$ strictly from faithful instances.
Unsupervised Variant: When oracle labels are unavailable, the method estimates $F_0$ directly from all calibration samples. Theoretical analysis (Proposition 1) proves that if the contamination rate of hallucinated samples is low, the estimated CDF remains robust.
Theoretical Guarantees: The authors generalize the Dvoretzky–Kiefer–Wolfowitz (DKW) inequality to random-length vectors to provide finite-sample calibration guarantees. They further prove that the false positive and false negative rates of the CES test decay exponentially with generation length.

Key Contributions

Entropy Distributions as Fingerprints: The paper provides empirical and theoretical evidence that the shape of the token entropy distribution (beyond the mean) distinguishes hallucinated from faithful generations. A synthetic mean-shift experiment confirms that shape differences persist even after removing location shifts.
Statistically Principled Framework: CES formalizes detection as a hypothesis test, separating the definition of hallucination (delegated to an oracle) from the detection mechanism. This makes the method agnostic to the specific model, task, or hallucination taxonomy.
Theoretical Guarantees: The work establishes exponential convergence of error rates with generation length and provides finite-sample bounds via a novel random-length DKW inequality.
Extensive Empirical Validation: CES was evaluated across 8 QA benchmarks and 10 models (spanning open-source Llama/Falcon/Mistral families and proprietary GPT-4 variants).

Experimental Results

Performance: Across 80 model-dataset experiments, CES (unsupervised) achieved an average rank of 6.29 among 17 methods, placing it in the top statistical clique alongside complex multi-sample methods like Kernel Language Entropy (KLE).
Comparison to Baselines:
- CES significantly outperformed single-pass baselines like Perplexity and Length-Normalized Entropy.
- CES achieved a 66.8% win rate in pairwise comparisons against 16 benchmark methods.
- Remarkably, CES (single-pass, black-box) was statistically indistinguishable from multi-sample methods that require $K \ge 5$ forward passes, effectively closing the gap between lightweight and expensive detection.
Robustness: The method demonstrated resilience to calibration contamination (up to 50% hallucinated tokens in the reference set) and noisy oracle labels, with performance remaining stable even when 50% of calibration labels were flipped.
API Generalization: CES successfully transferred to API-only models (GPT-4 family) using only top-k log-probabilities, achieving comparable performance to open-weight models.

Significance and Claims

The paper claims that CES represents a significant advancement by providing formal error guarantees (Type I and Type II error bounds) that existing heuristics lack, while maintaining the computational efficiency of a single forward pass.

Practical Deployment: By requiring only black-box logit access and a single pass, CES is suitable for real-time, large-scale deployment in high-stakes settings where multi-pass methods are too costly and internal-state methods are inaccessible.
Theoretical Insight: The work shifts the paradigm from viewing entropy as a scalar uncertainty metric to treating it as a distributional fingerprint, demonstrating that the "tail" of the entropy distribution is a critical signal for hallucination.
Limitations: The authors acknowledge that CES power degrades for very short generations ( $m < 10$ tokens) due to insufficient samples for distributional testing. Furthermore, if a model is trained on incorrect data (systematic hallucination), the entropy distribution may not diverge, rendering the test powerless. The method also relies on access to next-token probabilities, which some API providers may restrict.

In conclusion, the paper argues that CES offers the strongest detection performance achievable under minimal-access constraints, providing a rigorous, statistically grounded alternative to both heuristic single-pass methods and computationally expensive multi-sample approaches.