MMD-Flagger: Leveraging Maximum Mean Discrepancy to Detect Hallucinations

arXiv:2506.01367 · cs.CL, stat.ML · Submitted 2026-08-19 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Next we'll be talking about the paper "MMD-Flagger: Leveraging Maximum Mean Discrepancy to Detect Hallucinations".

Jane: The paper was written by Kensuke Mitsuzawa and Damien Garreau from University Côte d’Azur and CNRS and Center for Artificial Intelligence and Data Science (CAIDAS) and Julius-Maximilans-Universität Würzburg, Germany.

Tom: Stay tuned as we take you through the paper and discuss its implications.

Paper discussion segment 1 — Tom and Jane discuss title and authors of the paper 'MMD-Flagger: Leveraging Maximum Mean Discrepancy to Detect Hallucinations' and its implications. Explain in simple terms; do not repeat what earlier segments covered.: Tom: We’ve seen how MMD-Flagger is designed to measure statistical divergence, but let’s look at the authors’ specific approach now that they are presenting the core findings of this work.

Jane: The central idea is fascinating because it relies on generating a 'hypothesis' output using a standard decoding strategy, and then comparing it to what happens when you vary the temperature.

Lu: It’s about creating a dynamic comparison; instead of just looking at one sample, we’re observing the entire *trajectory* of similarity as the temperature changes.

Meng: That requires running multiple stochastic samplings for every single input, which is a significant computational load that makes sense only if the resulting data is so valuable.

Lalam: I see this as a huge shift in perspective, Lalam; we are treating language generation not as a static sequence, but as an output derived from an underlying probability distribution that needs validation.

Tom: It’s clear they found that when the system is generating something truthful, the outputs tend to cluster together across different temperature settings.

Jane: But when a hallucination occurs, those stochastic samples start spreading out or diverging significantly from the initial hypothesis.

Lu: This divergence is where Maximum Mean Discrepancy comes in—it measures that increasing distance between two distributions as you move away from zero temperature.

Meng: The practical implication is that they are essentially using randomness to expose the weakness of the deterministic generation process, which is a clever way to test robustness.

Lalam: This approach suggests that AI isn't just 'making things up'; it's exhibiting a predictable statistical failure mode, one that can be located and measured using established mathematical frameworks.

Paper discussion segment 2 — Tom and Jane discuss the paper's summary of the paper 'MMD-Flagger: Leveraging Maximum Mean Discrepancy to Detect Hallucinations' and its implications. Explain in simple terms; do not repeat what earlier segments covered.: Tom: Following our discussion of how MMD-Flagger works, let’s zero in on the specific mechanism they use for detection—that distinctive U-shaped trajectory.

Jane: The authors discovered that the signature of a hallucination is a U-shaped curve when you plot the similarity between the default output and your stochastic samples as temperature changes.

Lu: That’s a huge leap in concept; it means the system is identifying an inflection point where its own probabilistic behavior breaks down, which is far more complex than just checking if tokens match.

Meng: It seems like they are looking for a specific minimum point in the MMD curve, tau min, and then checking if that minimum temperature is high enough to trigger their detection threshold tau zero.

Lalam: This allows us to see the expected level of variance based on context clues; if the model starts behaving randomly at low temperatures, it' not grounded in factual data.

Tom: It’s interesting how they define that "U-shape"—it’s a visual way of saying that the randomness only becomes significantly different from the truth in certain temperature ranges.

Jane: Exactly, Tom; it suggests that if the model is hallucinating, its output is unstable across those intermediate temperatures, which is a key insight.

Lu: This moves us past subjective assessments of 'sounding convincing' and into objective, quantifiable metrics of distributional fitness.

Meng: So, from an implementation standpoint, we are relying on observing this specific shape in the MMD graph to move away from simply having a fixed threshold that might flag too many correct answers.

Lalam: Knowing that this method has such a solid theoretical backing gives us confidence that it can be adapted and scaled for various industries where factual accuracy is paramount.

Paper discussion segment 3 — Tom and Jane discuss the paper's summary of the paper 'MMD-Flagger: Leveraging Maximum Mean Discrepancy to Detect Hallucinations' and its implications. Explain in simple terms; do not repeat what earlier segments covered.: Tom: We’ve established that MMD-Flagger identifies hallucinations using a U-shaped trajectory, but now let’s look at the technical refinements—the practical upgrades they suggest for making this robust.

Jane: The authors point out that we shouldn't just rely on one single discrepancy threshold; they advocate for 'adaptive flagging' based on how reliable the optimal candidate tokens are.

Lu: That’s a massive leap in concept; it means the system should assess its expected level of variance depending on whether the AI is summarizing facts or generating creative concepts.

Meng: So, if the AI is summarizing established facts, we can expect near-zero deviation, but when it's extrapolating, our tolerance for deviation needs to be higher?

Jane: That’s right. But if it’s generating something creative, a small degree of educated conjecture might actually be permissible within that specific domain.

Tom: That concept of dynamic tolerance is huge; it treats AI detection not as a simple pass/fail switch but as a continuous gradient assessment of how far the output is from the expected pattern.

Lu: Precisely, and the statistical profile of a truthful generation is expected to be tightly clustered, so any signal suggesting a wider spread indicates a significant probabilistic breakdown.

Meng: This leads to an implementation question: when we need highly optimized kernel calculations running on specialized hardware, how does this framework scale up beyond these controlled benchmarks?

Lalam: It suggests that we're moving toward verifying the structural integrity of knowledge itself, Lalam; assessing if the AI is merely mimicking a pattern or actually grounded in a real concept.

Jane: That’s right. They are defining 'expected' based on how reliably a clear optimal token exists in that domain for every single prompt.

Tom: It elevates the entire concept of verification from simple fact-checking to a full comparative analysis of informational coherence across different temperature settings.

Lu: I find the methodology fascinating because we’re not comparing two words, we are mapping the entire distribution of potential outputs into a mathematical space to see if it aligns with the ground truth.

Meng: From an operational standpoint, this means they are taking the original input and running it through an LLM to get a 'hypothesis' output, then generating twenty-five alternative versions using various temperature settings.

Lalam: It suggests that we are assessing the probability landscape of language; if the AI isn't converging on a statistically reliable path, it’s not just random noise—it’s structurally devi from expected factual data.

Conclusion: Tom: We have covered how MMD-Flagger improves detection and its potential for general semantic verification; it’s truly a major breakthrough in the field.

Jane: It’s clear that this method allows us to move toward a level of verification that was previously unimaginable in generative models, fundamentally changing the trust dynamic around AI output.

Lu: I think the most important thing for us to take away from this research is that our understanding of AI’s limitations has deepened, showing us a measurable a way to assess how much its output diverges from expected patterns.

Meng: The fact that we can see competitive performance across diverse datasets suggests the operational cost is manageable for real-scale deployment, which is a huge relief for my team because scalability is always paramount.

Lalam: It’s an amazing thing to see, because as we move toward using this technology in global contexts, the reliability offered by MMD-Flagger promises more trust and less confusion for everyone involved in the field.

Tom: I agree with Lalam; it gives us a powerful way to say that we're not just accepting whatever the AI spits out but demanding a quantifiable, measurable level of factual consistency.

Jane: It’s definitely encouraging to finish this discussion by acknowledging that MMD-Flagger: Leveraging Maximum Mean Discrepancy to Detect Hallucinations is a sophisticated tool, not just some quick fix for hallucinations—it's foundational.

Lu: It offers such an elegant framework for understanding the subtle ways in which generative models might fail to capture grounded knowledge, moving beyond simple syntax checks entirely.

Meng: And I think it provides a very solid baseline for future work that needs to build upon this foundational statistical method; it gives us a clear roadmap forward.

Lalam: The promise of MMD-Flagger is that it can help us correct those specific instances where AI says something that sounds fluent but isn't factually sound, ensuring a much better experience for the end-user.

Kensuke Mitsuzawa, Damien Garreau

University Côte d’Azur · CNRS · Center for Artificial Intelligence and Data Science (CAIDAS) · Julius-Maximilans-Universität Würzburg, Germany

cs.CL, stat.ML

Submitted: 2026-08-19

Updated: 2026-08-20

Code: https://github.com/amazon-science/abstractive-factual-tradeoff

Importance score: 78/100

The gist: MMD-Flagger: Leveraging Maximum Mean Discrepancy to Detect Hallucinations Introduction and Motivation Large language models (LLMs) are widely used in everyday life, but a "fundamental obstacle

Key concepts

Maximum Mean Discrepancy (MMD)
MMD is a mathematical framework used to measure the statistical divergence between two distributions. In this context, it measures the increasing distance between outputs generated by an AI model at different temperature settings, helping to quantify how much of its probabilistic behavior is changing.
U-shaped Trajectory
This signature is identified when plotting the similarity between a default AI output and stochastic samples as temperature changes. A U-shaped curve indicates that the AI's probabilistic behavior is unstable or diverging, serving as a quantifiable metric for detecting hallucinations.
Stochastic Sampling at Varying Temperatures
The method involves running multiple random samplings (stochastic samples) for every input while varying the temperature setting. This process allows researchers to observe the entire trajectory of similarity, revealing whether the AI's output clusters reliably or spreads out.

Terminology

Summary

MMD-Flagger: Leveraging Maximum Mean Discrepancy to Detect Hallucinations

Introduction and Motivation

Large language models (LLMs) are widely used in everyday life, but a fundamental obstacle prevents their use in many critical applications: their propensity to generate fluent, human-quality content that is not grounded in reality. The detection of such hallucinations is therefore of the highest importance. This paper addresses the challenge of detecting hallucinations by proposing MMD-Flagger, a novel method based on Maximum Mean Discrepancy (MMD), which measures a non-parametric distance between distributions.

The emergence of hallucinations is linked to the decoding strategy used for next-token prediction. While beam search is often practical, it is prone to the so-called likelihood trap... where sequences with high probability can nevertheless yield poor-quality or uninformative outputs, which can cause the hallucination. The authors propose an alternative: stochastic sampling.

The Core Concept of MMD-Flagger

The central idea of this study is to inspect a hypothesis output (y hyp) generated with a default decoding strategy and compare it against stochastic samples obtained under varying temperature (tau).

  1. Generation: For a given input x, the hypothesis output y hyp = LLM hyp(x) is obtained.

  2. Stochastic Sampling: Additional documents, Y sto, are generated by stochastic sampling at various temperature parameters tau.

  3. MMD Calculation: The MMD between y hyp and the new groups (Y sto) is computed for each tau. This creates a trajectory of MMD values corresponding to different temperature settings.

The authors observe that at low temperature settings, we observe that the sampled outputs are highly similar to the hypothesis if the model is not hallucinating, and moderately dissimilar if it is. This behavior occurs because when a clear optimal candidate token exists, the default decoding strategy and stochastic sampling tend to favor the most probable token.

Detection Mechanism

The method relies on inspecting the shape of this MMD trajectory. By plotting the similarity between the default output and stochastic samples as a function of temperature, two distinct curves emerge:

  • A monotonously increasing trajectory in the nonhallucinating scenario.

  • A U-shape curve in the hallucinating scenario.

The authors define a hallucination by checking if tau min (the temperature parameter realizing the minimal MMD distance) is greater than or equal to a minimum threshold tau 0. This is visualized in Figure 1, where we flag y hyp as a hallucination when the minimum value of MMD falls within the blue region.

Methodological Details

  • MMD Definition: MMD measures distance by mapping distributions into a Reproducing Kernel Hilbert Space (RKHS). A large MMD value indicates significant differences between two distributions.

  • Vector Representation: To capture semantic relationships, sequences are converted into dense token embedding vectors. Two aggregation methods are used: 'avg' (computing the average over T vectors) and 'concat' (joining the T embeddings to form a fixed-length vector).

  • Kernel Choice: The study utilizes the linear (Dot) and Gaussian kernels, with specific parameters like gamma or percentile values chosen based on calibration data.

Empirical Assessment and Results

MMD-Flagger was benchmarked on four datasets: machine translation benchmarks (LFAN-HALL, Halomi) and summarization datasets (ConstFact, FaithBench). The results are reported using recall and precision.

  • In the LFAN-HALL dataset, MMD-Flagger (avg-Gaussian) achieved a recall of 0.808 (Precision: 0.362).

  • The method's performance was found to be competitive relative to natural competitors.

*When comparing average and concatenation aggregation, the 'concat' aggregation generally yielded slightly superior or comparable performance to avg.

Limitations and Discussion

The authors acknowledge several limitations:

  1. Computational Cost: MMD-Flagger requires repeated stochastic sampling across multiple temperature settings, which scales with the size of LLM and the length of the generated sequence.

  2. Argmax Decoding: The method is not intended to inspect hypothesis outputs generated via argmax selection because argmax produces rigid sequences lacking the variability necessary for MMD-Flagger's stochastic distribution analysis.

Conclusion

The authors conclude that MMD-Flagger achieves detection performance comparable to the strongest uncertainty-based estimation methods on two machine translation benchmarks. The method's performance is robust with respect to changes in the output vectorization and choice of the kernel.

Improvements for AI systems

A dedicated post-processing module, MMD-Flagger, will be integrated into the LLM pipeline immediately following generation (post- y hyp and Y sto generation).

  • Mechanism: The system captures the initial deterministic output (y hyp) and runs N stochastic samples (Y sto) across a defined temperature range tau = 0.1,, 1.0. It then computes the Maximum Mean Discrepancy (MMD) between y hyp and Y sto.

  • Specific Enhancement: The system tracks the resulting MMD trajectory across tau. Instead of a fixed similarity threshold, it dynamically identifies the temperature parameter (tau min) that minimizes the MMD distance.

  • Detection Logic: If tau min is found to be greater than or equal to a calibrated minimum threshold (tau 0), the output is flagged as an LLM Hallucination. Otherwise, it is deemed plausible.

The system will utilize dual vector aggregation methods to balance computational overhead with detection precision:

  • Average Aggregation (Default/High Efficiency): The system computes the average of all token embedding vectors (H) for the sequence. This provides a fixed-length, robust representation, minimizing computational complexity and ensuring stable performance, especially useful for real-time applications.

  • Concatenation Aggregation (High Sensitivity): For high-stakes environments where precision is paramount, the the system concatenates all token embeddings (H). This preserves fine-grained semantic detail but requires dynamic resource allocation based on sequence length.

MMD-Flagger principles can be repurposed during model development:

  • Pre-Training Divergence Measurement: The system can use MMD to measure the distributional shift between the expected output distribution (from a ground truth dataset) and the actual output distribution of a model, allowing developers to identify specific training phases where divergence (potential hallucination) begins.

  • Hallucination-Aware Fine-Tuning: During alignment, MMD can be used as a loss function component to penalize model outputs that exhibit high variance or U-shaped MMD trajectories when stochastic samples are generated against the reference data.

The improved system provides the following specific capabilities:

  1. Automated Hallucination Filtering: It can autonomously scan large volumes of LLM-generated content (e.g., summarizing news, translating documents) and flag outputs with high probability as hallucinations without human intervention.

  2. Self-Contained Auditing: Unlike existing methods that require external models (e.g., BERT, T5), this system operates entirely on the internal representations of the generating LLM, providing a fast, resource-efficient auditing capability for internal quality assurance pipelines.

  3. Robust Error Classification: It can differentiate between true hallucinations (U-shaped MMD trajectory) and syntactically plausible but semantically incorrect outputs (e.g., simple translation errors), reducing false positives inherent in simpler similarity metrics.

  4. Adaptive Performance: The system is designed to handle diverse datasets by automatically calibrating the tau min based on the dataset's unique hallucination profile, ensuring high recall and precision across different domains (e.g., machine translation vs. abstractive summarization).

Sources

Related papers