Imagine training a deep generative model on a complex dataset of natural images, such as CIFAR-10. Intuitively, you might expect the model to assign high probability density to images resembling the training set and low density to completely unrelated images, such as simple geometric shapes or blank canvases.
However, as Eric Nalisnick et al. demonstrated in Do Deep Generative Models Know What They Don't Know?, the opposite frequently occurs. High-capacity generative models regularly assign higher likelihoods to out-of-distribution (OOD) data than to their own training data. This counterintuitive phenomenon is known as the likelihood paradox.
Understanding why this happens requires moving past the assumption that a high probability density value ($p(x)$) implies that a sample is "typical" of a distribution. Instead, this failure mode stems from a combination of high-dimensional geometry and an inherent architectural bias toward structural simplicity.
Density Versus Mass in High Dimensions
To build intuition for the paradox, we must look at how probability behaves in high-dimensional spaces. In a low-dimensional space, the mode (the point of highest density) sits comfortably in the region where most of the data resides. But as the number of dimensions grows, the volume of space expands exponentially away from the center.
As explained by Andi Zhang et al. in Falsehoods that ML researchers believe about OOD detection, the majority of a distribution's probability mass in high dimensions concentrates not at the mode, but in a thin spherical shell far from it, known as the typical set. This behavior is a direct consequence of the Gaussian Annulus Theorem, which Henry B. Moss et al. highlight in Return of the Latent Space COWBOYS: Re-thinking the use of VAEs for Bayesian Optimisation of Structured Spaces.
Because of this geometric property, an external, out-of-distribution dataset with lower variance can sit entirely inside the high-density region near the mode of the training distribution. The model assigns these OOD samples high point-wise density ($p(x)$) even though they are completely unrepresentative of the typical samples the model would actually generate. Consequently, raw density thresholds fail because they confuse high point density with membership in the typical set.
The Complexity Bias
The second reason for the paradox is that deep generative models are fundamentally biased toward low-level statistical regularities rather than high-level semantic concepts.
As Joan Serrà et al. show in Input Complexity and Out-of-distribution Detection with Likelihood-based Generative Models, a model's density estimation is heavily confounded by the description length or intrinsic complexity of the input. Simple, low-entropy inputs (like solid colors or repetitive patterns) naturally receive high likelihood scores simply because they are mathematically easier to compress and represent.
This means a model trained on highly complex data will perceive simpler OOD data as highly probable prototypes. Alireza Shafaei et al. observe in Does Your Model Know the Digit 6 Is Not a Cat? A Less Biased Evaluation of "Outlier" Detectors that generative models capture low-level statistical regularities rather than semantic meaning. Similarly, M. M. Amaan Valiuddin et al. demonstrate in Efficient Out-of-Distribution Detection of Melanoma with Wavelet-based Normalizing Flows that normalizing flows prioritize low-level pixel correlations and textures over semantic content. This focus causes them to penalize visually complex in-distribution images while rewarding simpler anomalies.
Overcoming the Paradox
To make generative models reliable for anomaly detection, researchers have developed methods to isolate distribution-specific fit from generic structural simplicity.
One prominent approach is to explicitly correct for input complexity. Joan Serrà et al. propose an OOD score that subtracts an estimate of the input's complexity from its negative log-likelihood, effectively leveling the playing field for complex data.
Another strategy focuses on how the model behaves when forced to adapt to a specific sample. The authors of Likelihood Regret: An Out-of-Distribution Detection Score For Variational Auto-encoder introduce "Likelihood Regret," which measures how much a Variational Autoencoder's (VAE) likelihood score improves when its latent representation is optimized specifically for a given test input. Because the model's information bottleneck restricts it from representing OOD data well without optimization, this delta serves as a robust anomaly signal.
Alternatively, instead of evaluating individual points, Dihong Jiang et al. suggest in Revisiting flow generative models for Out-of-distribution detection comparing the distribution of a group of samples within the model's latent space using statistical tests, bypassing point-wise likelihood pitfalls entirely.