Representation collapse occurs when internal representations in a large language model (LLM) lose their diversity and become nearly identical. This often results from deep architectures mapping diverse input patterns to a narrow set of output activations. As a result, the model's latent space—the high-dimensional space where inputs are mapped to vectors—shrinks, causing different inputs to converge to similar or even identical points. This loss of diversity limits the model's ability to capture and express the variety of meanings in language.

Key causes of representation collapse include:

• Excessive regularization;
• Overly aggressive normalization;
• Architectural choices that push activations toward a mean value.

Strong constraints to prevent overfitting can suppress features needed to distinguish inputs. Repeated normalization or certain activation functions can also cause activations to converge, erasing input distinctions.

Collapse reduces model expressivity. When latent space contracts, the model cannot differentiate subtle semantic or syntactic differences, leading to generic and less informative outputs. In severe cases, the model produces the same response for all inputs, signaling a complete loss of meaningful structure.

Geometrically, collapse is a loss of variance in the latent space. A healthy model has vectors pointing in many directions, encoding diverse features. Collapse causes vectors to cluster, reducing the spread (variance) and eliminating directions for separating concepts.

Oversmoothing is closely related. It describes representations becoming increasingly similar across layers, especially in deep networks. Oversmoothing erases differences between inputs, making final representations nearly indistinguishable. Both collapse and oversmoothing diminish the model's capacity to encode complex information, resulting in degraded performance and less interpretable latent structures.

Recognizing these issues is essential for designing and training LLMs that maintain rich and robust latent spaces.

Key Insights

• Representation collapse reduces the diversity of internal representations, making it difficult for the model to distinguish between different inputs;
• Oversmoothing causes representations to become increasingly similar across layers, compounding the effects of collapse;
• Loss of variance in the latent space signals that meaningful directions for encoding information are disappearing;
• Excessive regularization, normalization, or poor architectural choices are common causes of collapse and oversmoothing;
• The primary danger is a loss of model expressivity, leading to generic or uninformative outputs.